Program

Background: Circadian rhythms regulate diverse aspects of physiology, yet most fMRI studies rarely consider effects of time of day on measures of brain activity. Here, we examined the impact of time of day on resting state fMRI (rs-fMRI) in over 900 subjects scanned between 8am to 10pm over two sessions. Studies show that the magnitude of the brain’s global signal (GS) fluctuation varies with arousal state. For example, low arousal (e.g. following insufficient sleep) has been associated with higher, while high arousal (e.g. after coffee) has been associated with lower levels of GS fluctuation. Thus, based on known circadian variation in arousal levels, we hypothesised that GS fluctuation would be lowest in the morning, increase in the midafternoon, and dip in the early evening.

Methods: We used ICA-FIX denoised rs-fMRI scans from the S1200 release of the HCP dataset. Data was analysed from two sessions (Session 1: n=942; Session 2: n=869). GS fluctuation was defined as the temporal standard deviation (SD) of the global signal time-series in each run. Regional BOLD signal fluctuation and RSFC were computed from mean time-series in 419 regions of interest consisting of 400 regions from the Schaefer cortical parcellation and 19 subcortical regions derived using Freesurfer. Respiratory and pulse oximetry data were also analysed. Respiratory variation (RV) was defined as the standard deviation of respiratory waveform within 5.76-second windows. Effects of time of day were assessed by Pearson correlation of each variable of interest with time of day of scans across subjects. Significance was determined via permutation testing (k=100,000), during which family structure of participants was kept intact. Correlation matrices encoding RSFC were corrected for multiple comparisons using network-based statistic with an initial threshold of p<0.001. All significant results shown survived FDR correction at q<0.05.

Results: In contrast to our hypothesis, we observed time of day-associated reductions in GS fluctuation. RV SD (standard deviation of RV) also showed time of day-associated decreases, although effects of time of day on GS fluctuation remained significant after correcting for RV SD across participants. We also found negative correlations between time of day and regional BOLD signal fluctuation and between time of day and RSFC, particularly in visual and somatosensory cortices. The absolute magnitude of time of day-associated reduction in RSFC was greater than the association between RSFC and fluid intelligence in the same sample. All of these findings were replicated in both sessions.

Conclusion: We report robust time of day-associated reductions in the strength of resting state GS fluctuation, regional BOLD signal fluctuation and RSFC. These effects are not easily accounted for by expected patterns of diurnal variation in arousal levels or physiological artefacts. We recommend reporting and considering effects of time of day in future studies.

High-resolution fMRI is particularly useful for resolving fine details tangential to the surface as well as differences across cortical depth. Whereas most cortical-depth-dependent fMRI studies focus on measuring the degree of activation at different depths by averaging all the responses at each depth, there has been less work assessing the consistency of the spatial distribution of responses across depth. To address this, we performed a depth-dependent analysis of bottom-up and top-down driven somatotopic digit maps in human primary somatosensory cortex, S1.

Two types of somatotopic fingertip maps were obtained, driven primarily by (1) bottom-up or (2) top-down processes. The bottom-up maps were generated via a sensory condition in which phase-encoded vibrotactile stimulation was delivered across fingertips (index, middle, ring and little). The top-down maps were generated by sweeping attention across the fingertips with the same timing as in the sensory condition, while all four fingertips were under constant sensory stimulation.

Data were acquired on a Siemans Magnetom 7T scanner with a 32-channel head coil (NOVA Medical). An MP2RAGE sequence was used to collect whole-brain anatomical images (0.5mm isotropic). Functional data were collected using a 3D-EPI sequence (TR/TE=82/30ms, 0.8mm isotropic) positioned to cover S1 in the left hemisphere (contralateral to stimulated/attended fingers).

Anatomical datasets were used to guide an equivolumetric layering approach to define a precise set of laminar surfaces, onto which functional data were interpolated and smoothed tangential to the surface. A response delay analysis was then performed for both experimental conditions to yield a response delay value which can be used to estimate the preferred fingertip for every vertex in S1.

Both sets of somatotopic maps were found to vary across depth and were marked by a banded pattern at superficial layers; however, this pattern appeared to dissipate in deeper layers. We also assessed the consistency of the cortical fingertip representations across cortical depth. Based on cortical magnification, we expected a non-uniform distribution across the delay values, with more vertices preferring the index finger. We see this pattern of non-uniformity at the pial layer over both experimental conditions. However, the distributions tend to flatten out at deeper laminar depths – as suggested by the dissipation of the banding pattern.

Our findings in S1 are in line with previous work in visual cortex examining how 3D-EPI responses vary across depth (strongest and most spatially spread responses are found superficially). While superficial depths do display the strongest signal, this signal is also likely blurred by the presence of large surface veins. Since individual digit bands are still visible in the mid-layer delay maps, it may be beneficial to use mid-depth responses if the goal is to concisely map the digit representations in S1.

INTRODUCTION:

Over the past few years there has been an effort to generalize deep neural networks to non-Euclidean spaces such as surfaces and graphs – with these techniques collectively being referred to as geometric deep learning1. Here we demonstrate the power of these algorithms by using them, along with neuroimaging data, to predict brain function from anatomy in human visual cortex. The human visual hierarchy is comprised of a number of different cortical visual areas, nearly all of which are organized retinotopically2. That is, the spatial organization of the retina is maintained and reflected in each of these cortical visual areas. This retinotopic mapping is known to be similar across individuals; however, considerable inter-subject variation does exist, and this variation has been shown to be directly related to variability in cortical folding patterns and other anatomical features3,4. It was our aim, therefore, to develop a neural network capable of learning the complex relationship between the functional organization of visual cortex and the underlying anatomy.

METHODS:

To build our geometric deep learning model, we used the most comprehensive retinotopy dataset openly available – that from the Human Connectome Connectome Project5. This dataset includes 7T fMRI retinotopic mapping data of 181 participants along with their anatomical data represented on a cortical surface model. The data serving as input to our neural network included curvature and myelin values as well as the connectivity among vertices forming the cortical surface and their spatial disposition. The output of the network was the retinotopic mapping value (i.e., polar angle or eccentricity) for each vertex of the cortical surface model.

Developing the deep learning model involved three main steps: (1) training the neural network, (2) hyperparameter tuning, and (3) testing the model. During training, the network

learned the correspondence between the retinotopic maps and the anatomical features by exposing the network to each example in a training dataset. Model hyperparameters were then tuned by inspecting model performance using a development dataset. Finally, once the final model was selected, the network was tested by assessing the predicted maps for each individual in a test dataset (previously not seen by the network nor the researchers). Our final model included 12 spline-based convolution layers6, interleaved by batch normalization and dropout.

RESULTS AND DISCUSSION:

Here we demonstrate that our neural network accurately predicted not only the main features of both polar angle and eccentricity retinotopic maps but they were also able to predict nuanced variations in the retinotopic maps across individuals. More generally, this work demonstrates the potential of geometric deep learning to provide models able to predict individual differences in brain function from anatomy.

What is a sulcus? The cortical surface is made up of many convolutions, called gyri, delimited by folds, called sulci. The main sulci are considered as the limits between functionally and architecturally different regions. Additionally, cortex morphometry is used to quantify brain development and degenerative diseases.

Why is it essential to automate sulci recognition? Despite the many tools available for 3D visualization of sulci, sulci labelling is a long and fastidious process. It takes several hours for an expert to label all sulci in a single brain and reliable labelling requires the opinion of several experts. However, because of the large variability of the folding pattern in the general population, inferring developmental biomarkers requires the mining of data from a large number of brains. These biomarkers may correspond to characteristics of the sulci, such as size, depth or opening, which require the prior labelling of sulci. Therefore, automation of the sulcus recognition is essential.

Why is this a difficult task? Learning to label sulci is an extremely complex challenge for several reasons. First, sulci are highly variable structures, some sulci are even absent in more than 70% of brains and some subjects have up to 8 sulci missing. Additionally, each brain contains more than 120 different sulci and only a small number of segmentation algorithms are made for as many structures. Finally, the number of manually labelled subjects which can be used for supervised learning is limited (62 brains in this study).

Why use deep learning? Deep learning is revolutionising our daily lives. Concerning image analysis, convolutional neural networks (CNNs) have significantly boosted the performance of image classification algorithms in the main challenges of the web. Nevertheless, the training databases were gigantic. So what about neuroimaging? The problems posed do not always allow to have large databases, however many studies have shown impressive results even with a limited number of images.

Despite their current popularity, no CNNs-based approach has yet been proposed for cortical sulci recognition. In the OHBMx twitter conference, I will present the deep learning based model that I have implemented in the Morphologist toolbox, a widely used sulcus recognition toolbox included in the BrainVISA package (www.brainvisa.info). This model presents significantly better performance than the model previously implemented in the Morphologist toolbox and it is also much faster to apply.

Introduction: Most traditional asymmetry studies are based on population-level asymmetry but individual-level asymmetry is rarely investigated. Conventionally, analyses of brain asymmetry can be affected by specific vertices affected by noise and distortions owing to the registration process. Here, we describe cortical shape using the eigenspectrum of the Laplace–Beltrami operator, which decomposes the surface vertices across different spatial scales without the need for registration. We develop a shape asymmetry signature (SAS), which summarizes brain asymmetry across spatial scales. We show that individual’s SAS is a highly unique feature that allows accurate subject identification at relatively coarse spatial scales.
Method: We used longitudinal images from 200 healthy participants from the OASIS-3. OASIS-3 provides MRI images of the brain and surface meshes created by FreeSurfer. Following the framework of ShapeDNA (Reuter et al., 2006), we used Laplace-Beltrami eigenvalues as global geometry descriptors of the cortical surface meshes. These eigenvalues reflect spatial variation at discrete spatial scales, which are ordered from coarse (low eigenvalue indices) to fine spatial scales. After the eigenvalues were calculated, we subtracted the eigenvalue spectra for white surface of the left hemisphere from those for the right hemisphere in the same subject at each scale, which we denote as the SAS for that subject. We used the Euclidean distance to calculate the distances between the SAS from different images both within and between subjects from the first time point (T1) and a later time point (T2). We repeated the subject identification score for a range of eigenvalues needed and found the number to maximize the peak identification score. This number served as a meaningful way to truncate the number of eigenvalues that are needed to provide unique SAS properties, thus representing the relevant spatial scales of cortical asymmetry that uniquely identify individuals.
Results: The SAS can serve as a unique feature for subject identification. The peak identifiability is given by the combination of the SAS calculated with the first 66 eigenvalues, which represent low frequency spatial scales.
Conclusions: We demonstrated a novel descriptor of brain asymmetry. This SAS was unique for each subject, resulting in accurate identification of subjects. Surprisingly, the findings indicate that the uniqueness of asymmetry occurs at coarse spatial scales. Coupled with previous findings that neuropsychiatric disorders are related to abnormalities in asymmetry, our results suggest that such disorders have large underlying spatial structures that are not seen with point-wise methods focused on small morphological scales.

Mild traumatic brain injury, including sports-related concussions (SRC), accounts for 70-90% of all TBI cases. Recent evidence suggests that repeated SRC can lead to debilitating long-term neurological disturbances, including persistent post-concussion symptoms (i.e. lasting longer than one month). Despite increasing awareness surrounding the risks of SRC, the diagnosis and clinical management of athletes is largely guided by subjective and self-reported signs/symptoms which fail to reflect whether the brain has fully recovered and is no longer in a state of increased vulnerability. There is an urgent need for objective, diagnostic, and prognostic biomarkers that can better inform clinical decisions pertaining to SRC and return-to-play decision making. Here, we assessed the potential of advanced diffusion-weighted imaging (DWI) as a sensitive biomarker of SRC. Australian Football League (AFL) players with SRC (n = 8) were scanned using a 3T MRI at 48 hours post-injury and again at 2 weeks post-injury (i.e. at the time they were cleared to return to play). Their scans were analysed in comparison to a cohort of non-concussed AFL controls (n = 9) using MRtrix3. Fixel-based analysis of fibre density revealed robust differences between AFL controls and recently concussed AFL athletes at 48 hours post-injury. These differences were largely, but not completely, resolved by two weeks post-injury when concussed AFL athletes were no longer symptomatic. Although it is uncertain how these changes relate to underlying white matter pathology, previous studies suggest that DWI metrics may reflect axonal swelling, demyelination, and degeneration. As such, these results may be indicative of increased cerebral vulnerability that persists beyond the resolution of SRC symptoms. This has important implications for the clinical management of concussion, including guiding return to play decisions.

Traumatic brain injury (TBI) is particularly prevalent in the paediatric population (age 1-4 years). This is also an age when the brain is vulnerable to insult, in part due to the ongoing development of new white matter (WM) tracts. Following paediatric TBI, neurobehavioural deficits, emerge by adulthood, impacting the quality of life. Deficits seen are hypothesised to be symptomatic of disruption to the WM networks that are involved in the development of these behaviours. To understand the neuropathology of paediatric TBI, we used longitudinal in vivo diffusion weighted imaging (DWI) to characterise the extent of WM disruption after a mouse model of paediatric TBI, alongside a battery of behavioural tests to identify social deficits. Sham-surgery or a controlled cortical impact at two differing severities (mild or severe) was performed on male C57Bl/6 mice at postnatal day 21 (equivalent to a human toddler under 4 years of age), followed by acquisition of DWI scans at 4, 7, 28 and 70 days post injury. To test for social deficits, a battery of behavioural assessments were performed at approximately day 70 post injury. DWI metrics such as fractional anisotropy, axial and radial diffusivity were analysed to evaluate WM microstructural changes. Preliminary a-priori analysis showed loss of axonal integrity and myelin damage in WM regions such as corpus callosum, fimbria and the internal capsule, with evidence of evolving neuropathology over time. Deeper structures showed damage chronically while the subcortical WM regions showed damage more acutely. Severity-dependent changes in social behaviour were observed at adulthood. Future analyses will evaluate whether the observed neuroimaging findings are correlated with social dysfunction following paediatric TBI—with the aim of identifying predictive biomarkers to improve patient prognosis, and identify those at greatest risk of poor long-term outcomes.

Recent neuroimaging experiments have defined low-dimensional gradients of functional connectivity in the cerebral cortex that subserve a spectrum of capacities that span from sensation to cognition. Despite well-known anatomical connections to the cortex, the subcortical areas that support cortical functional organization have been relatively overlooked. One such structure is the thalamus, which maintains extensive anatomical and functional connections with the cerebral cortex across the cortical mantle. The thalamus has a heterogeneous cytoarchitecture, with at least two distinct cell classes that send differential projections to the cortex: granular- projecting ‘Core’ cells and supragranular-projecting ‘Matrix’ cells. Here we use high- resolution 7T resting-state fMRI data and the relative amount of two calcium-binding proteins, parvalbumin and calbindin, to infer the relative distribution of these two cell- types (Core and Matrix, respectively) in the thalamus. First, we demonstrate that thalamo-cortical connectivity recapitulates large-scale, low-dimensional connectivity gradients within the cerebral cortex. Next, we show that diffusely-projecting Matrix regions preferentially correlate with cortical regions with longer intrinsic fMRI timescales. We then show that the Core–Matrix architecture of the thalamus is important for understanding network topology in a manner that supports dynamic integration of signals distributed across the brain. Finally, we replicate our main results in a distinct 3T resting-state fMRI dataset. Linking molecular and functional neuroimaging data, our findings highlight the importance of the thalamic organization for understanding low-dimensional gradients of cortical connectivity.

Background: Psychotic disorders are associated with reductions in brain volume, but the timing and origin of these reductions remains unclear. In particular, the effects of antipsychotic medication and illness have been difficult to disentangle due to a lack of prospective, longitudinal, randomized placebo-controlled designs.

Methods: We conducted a triple-blind randomised placebo-controlled trial where 62 antipsychotic naïve patients with first episode psychosis (FEP) received either an atypical antipsychotic or a placebo pill over a treatment period of 6 months. Both FEP groups received intensive psychosocial therapy. A healthy control group was also recruited. Structural MRI scans were obtained at baseline, 3-months and 12-months. Our primary aim was to differentiate any illness-related brain volume changes from medication-related changes within the first 3 months of treatment. We secondarily investigated long-term effects at the 12-month timepoint.

Outcome: From baseline to 3 months, we observed a significant group x time interaction in the pallidum, such that patients receiving atypical antipsychotics showed increased volume, patients on placebo showed decreased volume, and healthy controls showed no change. In patients, a greater increase in pallidal grey matter volume over 3 months was associated with a greater reduction in symptom severity, consistent with a neuroprotective effect of atypical antipsychotics. We additionally found preliminary evidence for illness-related volume reductions in prefrontal cortices months follow-up and putative antipsychotic-related neurotoxicity in cerebellum at both 3-months and 12-months follow-up.

Interpretation: Our findings demonstrate that psychotic illness and antipsychotic exposure exert distinct and spatially distributed effects on brain volume. Our results also converge with prior work in suggesting that the therapeutic efficacy of antipsychotics may be primarily mediated through their effects on the basal ganglia.

Background: Neuroimaging studies have found altered connectivity of frontostriatal circuits across a broad spectrum of psychotic symptom severity. However, it is unknown whether dysconnectivity within frontostriatal circuits originates from disrupted bottom-up or top-down control signaling within these systems. We used dynamic causal modelling (DCM) to examine the effective connectivity of frontostriatal systems in first-episode psychosis (FEP).

Methods: A total of 52 FEP patients (26 males; mean [SD] age = 19.22 [2.92]) and 22 healthy controls (HCs; 14 males; mean [SD] age = 21.83 [1.8]) underwent resting-state functional magnetic resonance imaging. Stringent quality control was used to assess motion and physiological noise and preprocessing was implemented using FMRIPREP. Biologically plausible connections between eight regions in the left hemisphere encompassing the frontostriatal systems were modelled using spectral DCM. The regions comprise dorsolateral prefrontal cortex (DLPFC), ventromedial prefrontal cortex (VMPFC), anterior hippocampus, amygdala, dorsal caudate, nucleus accumbens, thalamus, and the midbrain. Differences in effective connectivity between patients and HCs was assessed using a parametric Bayesian model. Associations between effective connectivity parameters and psychosis symptoms, measured by the Brief Psychiatric Rating Scale, were assessed in the patient group in a separate Bayesian general linear model.

Results: Our results show that compared to HCs, patients demonstrate increased bottom-up effective connectivity from the midbrain to the anterior hippocampus (0.10 Hz) and to the nucleus accumbens (0.10 Hz). Simultaneously, patients showed reduced effective connectivity from the thalamus to the midbrain when compared to HCs (-0.10 Hz). Patients also demonstrated increased inhibitory self-connections of the DLPFC (0.09 Hz) and the midbrain (0.12 Hz); and reduced self-connection of the VMPFC (-0.12 Hz). Increased effective connections from the DLPFC, thalamus, and anterior hippocampus to the midbrain (0.08 – 0.12 Hz) as well as reduced effective connections from the midbrain to thalamus (-0.11) and from nucleus accumbens to hippocampus (-0.19 Hz) were associated with positive symptoms. We only report connections with strong evidence (Pp>0.95).

Conclusion: Increased effective connectivity from the midbrain to the hippocampus and nucleus accumbens in patients is consistent with the dopamine dysregulation hypothesis in psychosis and molecular imaging studies. Associations between positive symptoms and connectivity implicating the midbrain and the nucleus accumbens are also in line with elevated dopamine activity in psychosis, in which dopaminergic neurons in the midbrain is dysregulated due to the loss of inhibitory control from higher cortical areas, namely from the hippocampus. Here, we demonstrate that the results of our DCM are consistent with animal models of psychosis.

Experiments using naturalistic stimuli provide insights into cognitive function “in the wild”. However, challenges exist for creating parametric manipulations of such stimuli with tight experimental control. Here, we demonstrate how to selectively degrade subtle statistical dependences in natural scenes using the wavelet transform. Importantly, these manipulations leave basic features (e.g., luminance and contrast) intact.

To manipulate image structure using wavelets, the image is first decomposed using a family of wavelet basis functions sensitive to variance at specific spatial scales. We then randomly permute the decomposed data associated with one or more spatial scales – essentially destroying the structure at that scale. We conducted an fMRI experiment to demonstrate the application of these image manipulations for probing the functional architecture of the visual hierarchy. The overarching goal was to contrast levels of activity in different visual areas elicited by the presentation of intact vs. wavelet-degraded natural images.

Our results reveal a few salient response differences across visual area and experimental condition. Notably, we find evidence supporting our core hypothesis that higher cortical areas are more sensitive to the more complex statistical features of natural scenes (i.e., cortical areas respond more strongly when natural image structure is present than when absent and this difference increases as one progresses up the hierarchy). This work demonstrates the utility of using wavelet-based image manipulations to probe the visual hierarchy – supporting the notion that perceptual systems in the brain are optimally tuned to the complex statistical properties of the natural world.

Researchers have rightly highlighted the importance of anatomical precision in functional MRI (e.g. Devlin & Poldrack, 2007), yet if the aim is localizing function, then researchers need to actually test functional localization. Currently this is almost never done, imnplying that much of what we think we know about functional localization from fMRI might be wrong.

The current default method for “localizing” fMRI activation is unidirectional voxelwise thresholding of a statistical map. This approach implicitly accepts the null hypothesis that areas falling below threshold are not activated by the condition of interest. Accepting the null based on a unidirectional statistical threshold would not be accepted in other fields of science but remains the basis for localization claims in the vast majority of fMRI studies.

The problem is magnified by most fMRI studies being under-powered (Thirion et al., 2007), so that false negatives outnumber true positives. This has an alarming implication: The degree of apparent localization is greater (i.e. thresholded brain activation more focal) the more underpowered a study is! Given much higher statistical power, in studies with larger N for example, much more extensive, less focal regions of activation might be found. This is exactly what has been reported (Gonzalez-Castillo et al., 2012; Thyreau et al., 2012). Finding the same brain region activated over multiple experiments (e.g. in meta-analyses) offers little protection from false inferences of functional localization because the same regions of the brain might exceed threshold simply because they are consistently low noise regions (i.e. focal SNR hotspots).

Here I illustrate, with publicly shared fMRI data, explicit tests that can be used to generate Functional Localisation Maps (FLMs). In the first method, percent signal change values are extracted from suprathreshold clusters from a standard voxelwise analysis. Voxelwise permutation-based equivalence tests are then performed to identify those voxels elsewhere in the brain that show significantly less activation. A second method uses spatial mixture models that assign an explicit estimated probability that any one voxel belongs to the distributions of activated or non-activated voxels (Woolrich, Behrens, Beckmann, & Smith, 2005). The third method estimates voxelwise Bayes Factors, which provide evidence for both activation and no-activation for every voxel. Although imperfect, all offer better protection against false claims of functional localization than the traditional unidirectional thresholding approach.

Since the mid-1800s, language has been known to be lateralised in the brain, but the provenance and flexibility of hemispheric specialisation for language remain open questions. A wide variety of contributions to the field (psycholinguistic, neuropsychological, neuroimaging, and surgical mapping) has led to a range of findings and little consensus on interpretation. We tested language lateralisation and its plasticity in two fMRI experiments of intensive, ecologically-valid language training, using multi-factorial designs and a multi-pronged analytical approach. Converging patterns of results across experiments and analyses provide unique insights on hemispheric specialisation for language, and explain some conflicts in the field.

The study comprised a cross-sectional (n=34) and a longitudinal experiment (n=24). These were contrasted on participant language background and the language being studied, but tasks and analyses were common to both. In each experiment, participants carried out 2 tasks: a comprehension task and a production task. Standard preprocessing routines and level-I analyses were employed to obtain each subject’s whole brain activation. These were masked with broad neuroanatomical language network regions (inferior frontal gyrus, lateral temporal gyrus, inferior parietal cortex). Laterality indices were calculated for each task x language. Each subject’s native language (L1) lateralisation was used as a baseline for their non-native language (Ln), and learning-dependent changes in this relationship were examined in each modality:
(1) Group means and variation of laterality indices
(2a) Proficiency-dependent change (Cohen’s q) in L1-Ln correlation
(2b) Proficiency-dependent change (Cohen’s d) in same/opposite hemispheric dominance of languages

Results were consistent across experiments:
(1) Language production was left-lateralised, language comprehension was variable across individuals.
(2) Language comprehension exhibited substantial learning-dependent changes in hemispheric dominance, with languages tending to lateralise to opposite hemispheres with increasing proficiency. Production showed negligible change and remained left-lateralised.

Thus lateralisation for comprehension was seen to be variable and more flexible than for production. Findings explain comprehension-production asymmetry in language learning, and suggest involvement of sensorimotor systems in language lateralisation.

Takeaway: The comparative approach, multi-factorial design and within-subject analyses revealed robust principles of language lateralisation that were not apparent in more standard designs with group-level analyses and null hypothesis testing. We propose that a comparative approach and within-subject analyses are sensitive and robust in the face of well-known issues in neuroimaging — complicated interactions, small effect sizes, small sample sizes, and under-powered statistics — and produce replicable results even with modest sample sizes.

Most work in functional brain network characterisation assumes temporal stationarity of blood oxygen level dependent (BOLD) time series, and hence assumes that the brain remains in a fixed state throughout the experiment. However, this assumption is uncertain, since we know that the brain “waxes and wanes” exhibiting itinerant dynamics even when at rest. In this work, we present a novel Bayesian method for identifying how the brain shuttles between various states via model fitness assessment. Specifically, we detect network community change-point(s) based on overlapped sliding window applied to multivariate resting-state fMRI time series. We use the weighted stochastic block model to quantify the likelihood of a network configuration, and develop a novel scoring criterion that we call posterior predictive discrepancy by evaluating the goodness of fit between model and observations. The parameter space for this model includes latent labels assigning network nodes to interacting communities, and the block model parameters determining the weighted connectivity within and between communities. For inferring hidden variables, we use Monte Carlo methods based on Gibbs sampling to efficiently sample the posterior distribution over the parameter space.


The “replication crisis” in many scientific fields has raised concerns regarding the reliability of published results. One reason for the high rate of false positive results is the large number of “researcher degrees of freedom,” where the process of data analysis can be performed in multiple ways. This is specifically apparent in neuroimaging with a thriving “garden of forking analysis paths.” In the Neuroimaging Analysis Replication and Prediction Study (NARPS: www.narps.info/), we provide scientific evidence on the variability of neuroimaging results across analysis teams. We collected fMRI data from 108 participants performing two versions of the mixed gambles task, often used to study decision-making under risk. Seventy independent analysis teams received this dataset, including raw and preprocessed data. They freely analyzed it to test nine ex-ante hypotheses regarding activations in specific contrasts and brain regions based on previous results from studies using these tasks. The main outcome variable was measured as the fraction of teams reporting a significant result, based on their own criteria, for each hypothesis. We measured peer beliefs about the main outcome variable with two independent prediction markets (PMs): one for the members from the analysis teams, and one for expert scientists that did not participate in the analysis. The fraction of teams reporting significant results of the nine hypotheses varied from 6% to 84% (mean 28%). The rank correlation between the predictions and the main outcome variable was high for the "teams PM," and moderate for the "non-teams PM," but both groups substantially overestimated the support for the nine hypotheses. Based on the results of 70 independent analysis teams comprising 180 researchers worldwide, different choices of analysis pipelines have a substantial impact on neuroimaging results. Furthermore, we found that experts from the field are over-optimistic in predicting the support for neuroimaging hypotheses, even if they analyzed the data themselves. Our findings provide the first ecological scientific evidence on the variability of neuroimaging results across analysis methods in the wild and raise challenges on how to address variability of results due to researcher degrees of freedom.

To what extent electrocorticography (ECoG) and electroencephalography (scalp EEG) differ in their capability to locate sources of deep brain activity is far from evident. Compared to EEG, the spatial resolution and signal-to-noise ratio of ECoG is superior but its spatial coverage is more restricted, as is arguably the volume of tissue activity effectively measured from. Moreover, scalp EEG studies are providing evidence of locating activity from deep sources such as the hippocampus using high-density setups during quiet wakefulness. To address this question, we recorded a multimodal dataset from 4 patients with refractory epilepsy during quiet wakefulness. This data comprises simultaneous scalp, subdural and depth EEG electrode recordings. The latter was located in the hippocampus or insula and provided us with our “ground truth” for source localization of deep activity. We applied ICA for the purpose of separating the independent sources in theta, alpha and beta frequency band activity. In all patients subdural- and scalp EEG components were observed which had a significant zero-lag correlation with one or more contacts of the depth electrodes. Subsequent dipole modeling of the correlating components revealed dipole locations that were significantly closer to the depth electrodes compared to the dipole location of non-correlating components. These findings support the idea that components found in both recording modalities originate from neural activity in close proximity to the depth electrodes. Sources localized with subdural electrodes were ~70% more close to the depth electrode than sources localized with EEG with an absolute improvement of around ~2cm. In our opinion, this is not a considerable improvement in source localization accuracy given that, for clinical purposes, ECoG electrodes were implanted in close proximity to the depth electrodes. Furthermore, the ECoG grid attenuates the scalp EEG, due to the electrically isolating silastic sheets in which the ECoG electrodes are embedded. We conclude that in our current study, deep source localization accuracy of scalp EEG is comparable to that of ECoG.

With the advancement in artificial intelligence (AI) and machine learning (ML) techniques, researchers are striving towards employing these techniques for advancing clinical practice. One of the key objectives in healthcare is the early detection and prediction of disease to timely provide preventive interventions. This is especially the case for epilepsy, which is characterized by recurrent and unpredictable seizures. Patients can be relieved from the adverse consequences of epileptic seizures if it could somehow be predicted in advance. Despite decades of research, seizure prediction remains an unsolved problem. This is likely to remain at least partly because of the inadequate amount of data to resolve the problem. There have been exciting new developments in ML-based algorithms that have the potential to deliver a paradigm shift in the early and accurate prediction of epileptic seizures. Here we provide a comprehensive review of state-of-the-art ML techniques in early prediction of seizures using EEG signals. We will identify the gaps, challenges, and pitfalls in the current research and recommend future directions.

Introduction
Non-invasive neuromodulation with transcranial magnetic stimulation (TMS) has demonstrated moderate to good results and has been approved widely for clinical use. However, network studies have suggested that the brain is extremely difficult to control via a single site, which may contribute to the small effect size of the current single-coil techniques in therapeutic applications. The goal in the multinational ConnectToBrain project is to develop for research, diagnostics, and therapy a feedback-controlled multi-locus TMS (mTMS) technique.

Methods
The project includes three main development areas: 1) Multi-locus TMS array covering most of the cortical mantle and allowing real-time control over the locus, direction, intensity, and timing of the stimulation; 2) Real-time analysis of brain activity and connectivity by using high-density electroencephalography and a priori information from magnetic resonance imaging and magnetoencephalography for brain-state-dependent and closed-loop stimulation; 3) Demonstration of feasibility, safety, and effects of the developed techniques, and their therapeutic utility in dysfunctional brain networks in Alzheimer’s disease and motor stroke.

Results
We have performed feedback-guided motor-hotspot determination with a two-coil mTMS prototype, which is more reliable, less user-biased, and requires fewer pulses than the conventional manual approach. We are currently developing a five-coil prototype, and algorithms for brain-state-dependent stimulation.

Conclusion
We expect to develop new technology capable of correcting dysfunctional brain networks in several brain disorders with better therapeutic efficacy than current state-of-the-art techniques and the potential to induce a major paradigm shift in therapeutic neuromodulation. If successful and widely used in clinical applications, ConnectToBrain will eventually lead to a substantial reduction in the suffering and economic burden caused by brain disorders.

Decision-making is in the service of action regardless of whether the decision concerns perceptual information, goods or memories. Compared to recent advances in the neurobiology of perceptual or value-based decisions, however, the neural bases supporting the sampling of evidence in long-term memory, and the transformation of memory-based decisions into appropriate actions, are still poorly understood.
In the present fMRI study, we used multivariate pattern analyses (MVPA) to investigate the temporal dynamics of choice- and action-predictive signals during an item recognition task that manipulated the association between memory choices (old/new) and motor responses (eye/hand) across subjects. Choice-predictive activity was mainly observed in striatal, lateral prefrontal and lateral parietal regions, was sensitive to the amount of decision evidence and showed a rapid increase after stimulus onset, followed by a fast decay. Action-predictive signals were found in primary sensory-motor, premotor and occipito-parietal regions, were generally observed at the end of the decision phase and were not modulated by decision evidence. These findings suggest that a memory decision variable, potentially represented in a fronto-striato-parietal network, is not directly transformed into an action plan as often observed in perceptual decisions.
Regions exhibiting choice predictive activity, and especially the striatum, however, also showed a second peak of decision-related activity that, unlike pure choice- or actionpredictive signals, depended on the particular choice-response association. This second peak of activity in the striatum might represent the neural signature of the transformation of a memory decision into an appropriate motor response based on the specific choice-response association.

Unorganized heaps of analysis code are a growing liability as data analysis pipelines are getting longer and more complicated. This is worrying, as neuroscience papers are getting retracted due to programmer error. Furthermore, analysis code is increasingly published as the push towards open science continues, so the quality of your code becomes public knowledge.

I will present some guidelines that help keep analysis code well organized, easy to understand and convenient to work with:
1) Each analysis step is one script
2) A script either processes a single recording, or aggregates across recordings, never both
3) One master script to run the entire analysis
4) Save all intermediate results
5) Visualize all intermediate results
6) Each parameter and filename is defined only once
7) Distinguish files that are part of the official pipeline from other scripts

In addition to discussing the reasoning behind each guideline, an example analysis pipeline is presented as a case study to see how each guideline translates into code.

It is increasingly recognized that the sense of body ownership and body temperature are strictly interconnected. The temporary experimental induction of the sense of disownership for a body part in healthy subjects is accompanied by a decrease of body part temperature. On the other hand, in brain-damaged patients, the temporary remission from body part disownership (e.g., somatoparaphrenia) is complemented by increased temperature of both the arms. Individuals with Body Integrity Dysphoria (BID) may report an intense desire to have one of their healthy limbs to be amputated. We aim at exploring the hypothesis that this compulsive desire of amputation of an intact limb also derives from an impaired sense of ownership of this body part. To do this, we measured the limbs’ temperature of individuals with BID during a behavioural manipulation.
We administered a bodily awareness task to 7 individuals with BID seeking for amputation of one leg, and 7 healthy matched-controls. In a block-design experiment, participants were asked to direct their attention/awareness to one of their four limbs for 60s, with their eyes closed (awareness condition). Each trial of the awareness condition was followed by a rest trial in which participants were asked to try not to concentrate on any body part for 60s (rest condition). During the experiment, we recorded thermal image sequences of the surface of the subjects’ limbs using a 14-bit digital infrared thermal camera (AVIO, TVS-700, 320 9 240 Microbolometric Array).
We used Linear Mixed Models to study patterns of temperature changes in relation to the experimental conditions. Side (Left, Right), Limb (Arm, Leg), and Group (BID, Controls) were modelled as fixed factors. The delta value indicating the difference “awareness > rest” was computed for each limb and used as dependent variable. Random slope was modelled on Limb and random intercept was modelled on Subjects. Results showed a significant main effect of Limb (F(1,25)=7.818; p=.010)and an interaction Group by Limb (F(1,25)=12.269; p<.002). Post hoc Bonferroni-corrected comparisons demonstrated a significant bilateral decrease of legs temperature in BID compared to Controls (p=.003). Other main effects and interactions were not significant.
We demonstrated that when individuals with BID, with a desire for amputation involving the leg, focused their attention on the left and right legs, a decrease of temperature occurred. This effect is not present for the arms. Our findings may have potential therapeutic implications. In a Rubber Hand Illusion (RHI) experiment it has been demonstrated that increasing the temperature of the participants’ hand weakened the RHI. One might speculate that warming the legs of individuals with BID may result in a reduced desire for amputation. Future studies may explore such a hypothesis.

Introduction:
The autonomic nervous system (ANS) controls the majority of organ functions in the human body. It coordinates such central processes as circulation, digestion, metabolism and immune functions. The scientific investigation of acupuncture has long focused on pain, while some of its strongest effects (e.g. on nausea and vomiting, migraine, hypertension, and inflammation) are most likely mediated by the ANS. In this study, we measured the effect of electroacupuncture stimulation (EAS) on sympathetic and parasympathetic regulation using functional magnetic resonance imaging (fMRI) of the hypothalamus during specific autonomic tests.

Methods:
We compared EAS at two sets of acupuncture points that had previously shown strong (ST36-37) or weak (GB37-39) effects on autonomic function in animal studies. 15 healthy subjects were scanned twice on a 3T MR scanner. The protocol involved SMS-EPI functional scans (voxel size=2x2x2 mm³) and T1-weighted structural scans (voxel size=1x1x1 mm³). We conducted three different autonomic tests during the functional scans, each with a duration of five minutes: lower body negative pressure, isometric handgrip and slow breathing. Blood pressure and heart rate were measured concurrently to the functional scans. The entire block (3 autonomic tests, 1 rest run) was repeated once with concurrent electroacupuncture stimulation.
FMRI data were minimally preprocessed using tools from FMRIB Software Library (FSL v5.0.9) including motion correction, unwarping, temporal high-pass filtering and normalisation to a study template (Advanced Normalization Tools). We conducted a masked independent component analysis of the hypothalamus, excluding adjacent areas with high physiological noise. Afterwards, differences between EAS and no stimulation were calculated using dual regression and non-parametric paired t-tests. Results were corrected for multiple comparisons using threshold-free cluster enhancement and family-wise error correction.

Results:
When comparing the two acupuncture points, we found that ST36-37 had a significantly stronger effect on heart rate and blood pressure than GB37-39. This point-specific effect was consistent with a reduction of sympathetic excitability. Furthermore, we identified a set of hypothalamic nuclei mediating this effect, namely the paraventricular nucleus, the nucleus arcuatus and the supraoptic nucleus. We were also able to detect their connectivity changes to autonomic regions in the lower brainstem including the nucleus raphe pallidus and the ventral medulla.

Conclusion:
Our results underline that the ANS plays a central role in mediating acupuncture effects and that some of these effects are point specific.

Human odor perception relies on the tight interplay between the olfactory system and the trigeminal systems since most odorants stimulate both systems simultaneously, although to varying degrees. Interestingly, although not able to perceive the odorant itself, acquired anosmics are still aware of trigeminal components of odorants and often can use them to identify the odor. Additionally, odor perception is modulated by breathing patterns and pace as well as by expectancy, all of which also significantly affect neuronal activity. However, a sniff is not only part of the olfactory percept, but temporal changes in respiratory patterns have also been related to cognitive load and seem to be correlated with task-onset.
Following a brief overview of the central processing of the chemical senses, we will investigate the possibility to separate the different components of odor perception – respiration, trigeminal, and odorant effects – in normosmic and anosmic participants. To this end, we were using a pseudo-free breathing paradigm where participants had to performed a simple breathing experiment where participants either had to match their breathing cycle to an expanding or shrinking ball or were allowed to breathe at their own pace. During some of the inhale-phases olfactory stimuli (rose or peppermint), or diluted carbon dioxide (CO2) as pure trigeminal stimulus was presented. To control for expectancy and top-down processing, some of these presentations were preceded by a red cross to announce the upcoming stimulus but not its quality.
Results show that olfactory stimuli compared to guided breathing exhibited more robust and distinct activation within the piriform cortices but also within the insula, dorsolateral prefrontal cortices, and the parahippocampal and entorhinal cortices when the odor was announced. However, both effects were not observed in direct contrasts. In anosmic participants, a similar pattern evolved; however, at a very lean threshold. CO2, on the other side, activated somatosensory regions, yet comparing CO2 to breathing, revealed additional strong activations within the insular cortex and the supramarginal gyrus in all participants. Breathing itself activated expected areas within the basal ganglia, cerebellum, and frontal areas. As expected, temporal breathing patterns matched the presentation of an odorant only in guided condition, but not during free self-paced breathing.
These results further previous literature and emphasize the importance of the tight control of respiration patterns and cognition as an essential part of olfactory perception.

The mass-univariate analysis framework remains the most popular approach for task-fMRI inference. However, this method suffers from at least two fundamental limitations: First, with ample power, the smallest effects will always reach significance, meaning that application of the traditional approach to population-size neuroimaging datasets (e.g. UK Biobank, N=40,000) will result in essentially universal brain activation. Second, for any N, the standard procedure does not express the variability in the sizes and shapes of significant clusters that could be expected from repeated sampling of the population.
To overcome this, we developed Confidence Sets (CSs) on clusters found in thresholded maps. While statistical methods indicate where the null, i.e. an effect size of 0, can be rejected, the CSs are statements about where effect sizes have exceeded, and fallen short of, a non-zero threshold c, providing an upper CS (voxels with a true Cohen’s d effect size greater than c) and lower CS (voxels outside this set have a true Cohen’s d effect size less than c). The assertion is made with a (1 - a)100% confidence level, holding simultaneously for both regions.
The method is straightforward: by applying a bootstrap schema to specialized residuals, the upper CS is obtained by thresholding the estimated Cohen’s d field at a critical level slightly above c. Similarly, the lower CS is obtained by thresholding the estimated Cohen’s d field at a critical level slightly below c. The bootstrap methodology ensures that these thresholds are chosen to maintain the desired confidence level.
We assessed the performance of the method on four simulated signals representative of fMRI clusters, with homogeneous and heterogeneous noise. We considered a confidence level of 1 - a = 0.95 and sample sizes of N = 60, 120, 240 & 480. For a threshold of c = 0.8, we computed the ‘coverage’, the percentage of times that the true thresholded Cohen’s d was contained between the upper and lower CSs. Across all signal types and samples, empirical coverage hovered slightly above the nominal target, suggesting that the method can be effective even when the sample size is modest.
Finally, we applied the method to HCP working memory task data, operating on Cohen’s d effect size maps, where we obtained CSs for a variety of cluster forming thresholds. Here, the upper CSs localized activation in cognitive regions commonly associated with working memory, determining a Cohen’s d effect size of (at least) 0.5 with 95% confidence. By comparing the CSs with traditional thresholded statistical results, we observed how the CSs could provide an improved localization of practically significant effects.
In sum, the CSs synthesize information about the magnitude and reliability of effect sizes that is usually provided separately in an effect estimate and t-statistic map. For population neuroimaging studies, tools like this for interpreting the spatial profile of effects will become more important.

The brain can be modelled as a network with nodes and edges derived from a range of imaging modalities: the nodes correspond to spatially distinct regions and the edges to the interactions between them. Whole-brain connectivity studies typically seek to determine how network properties change with a given categorical phenotype such as age-group, disease condition or mental state. To do so reliably, it is necessary to determine the features of the connectivity structure that are common across a group of brain scans. Given the complex interdependencies inherent in network data, this is not a straightforward task. Some studies construct a group-representative network (GRN), ignoring individual differences, while other studies analyse networks for each individual independently, ignoring information that is shared across individuals. We propose a Bayesian framework based on exponential random graph models (ERGM) extended to multiple networks to characterise the distribution of a entire population of networks. Using resting-state fMRI data from the Cam-CAN project, a study on healthy ageing, we demonstrate how our method can be used to characterise and compare the brain's functional connectivity structure across a group of young individuals and a group of old individuals.

Making citizens active is one of the major challenges of the 21st century. At a time where issues of obesity are major societal concerns, scientific advances aim to find the keys to promote a healthier lifestyle. Engagement in physical activity on a long term basis is a complex phenomenon, which depends not only on the pleasure felt but also on the smooth functioning of the executive functions of the brain to avoid stopping when it gets too effortful. The objective here was to develop the use of fNIRS to reveal frontal involvement for inhibiting the urge to stop during moderate physical activity. fNIRS is a non-invasive imaging method that makes use of optical proprieties of light in order to evaluate local haemodynamic responses in a given cortical area. This tool is particularly salient to exercise-related protocols given that it is far less sensitive to movement than traditional scanning techniques such as fMRI or EEG. However, scientific community is sorely lacking of validation studies systematically comparing fNIRS suitability in simple vs. whole-body movements in order to confirm the use of this neuro-imaging technique to measure brain functions while moving. To address this methodological ellipsis, participants performed both a finger-tapping task (i.e., simple motor task) and a cycle ergometery task (i.e., complex motor task) in synchronisation with an external metronome, controlling the temporal complexity (i.e., slow vs. fast pace cuing). Changes in oxygenated haemoglobin concentrations in both frontal and motor areas were recorded. Results indicated that the cycle ergometery task was more prone to motor artefacts and physiological contamination compared to the finger-tapping task. Furthermore, larger shifts of the headset during whole-body movements were also noticed, which is problematic insofar as the optodes location varies across the task. Hence, this results provided strong support for the need to continue the technical development of protocol building (e.g., adding 3D measurement to have an online control on the headset positioning) and data-processing (e.g., rejecting frequency bands of physiological noise from the data) in order to be able to correctly use fNIRS during physical activity. Overall, this work allowed us to develop a strong data analysis pipeline to minimize the reported issues, notably by including participant and channel selection.

Structural and functional characteristics of the cortex systematically vary along global axes as a function of cytoarchitecture, gene expression, and connectivity. The topology of the cerebral cortex has been proposed to be a prerequisite for the emergence of human cognition and explain both the impact and progression of pathology. However, the neurogenetic origin of these organizational axes in humans remains incompletely understood. To address this gap in the literature our current study assessed macro scale cortical organization through an unsupervised machine learning analysis of cortical thickness covariance patterns and used converging methods to evaluate its genetic basis. In a large-scale sample of twins (n=899) we found structural covariance of thickness to be organized along both an anterior-to-posterior and inferior-to-superior axis. We found that both axes showed a high degree of correspondence in pairs of identical twins, suggesting a strong heritable component in humans. Furthermore, comparing these dimensions in macaques and humans highlighted similar organizational principles in both species demonstrating that these axes of cortical organization are phylogenetically conserved within primate species. Finally, we found that in both humans and macaques the inferior-superior dimension of cortical organization was aligned with the predictions of the dual-origin theory, highlighting the possibility that the macroscale organization of primate brain structure is subject to multiple distinct neurodevelopmental trajectories. Together, our study establishes the genetic basis of natural axes in the cerebral cortex along which structure is organized and so provides important insights into the organization of human cognition that will inform both our understanding of how structure guides function and for the progression of pathology in diseases.

The contrast of the interface between the neocortical grey matter and the white matter is emerging as an important neuroimaging phenotype for several brain disorders. In autism spectrum disorder (ASD), the study of Andrews et al (2017) showed a significant decrease of this contrast in several areas. We tried to replicate this study in two large cohorts: ABIDE 1 and 2 (Autism Brain Imaging Data Exchange, N=1,477) and EU-AIMS (European Autism Interventions – A Multicentre Study for Developing New Medications, N=586).

We did find a statistically significant result in the ABIDE cohort, although the direction of the effect was the opposite: instead of a reduction we found an increase of grey-white matter contrast in the ASD group. We then tried to replicate the effect in the EU-AIMS cohort, but we did not find any significant effect. Upon closer examination of the ABIDE analysis, we made a concerning observation: the statistically significant effect was due only to 1 site: NYU (New York University). Analysing the NYU site separately produced the same effect we observed in the complete ABIDE cohort, but stronger. Everything suggests that the only source of our statistically significant finding is due to a particularity of the NYU data. We tried to control for a difference in the amount of motion between the ASD and control groups, using motion parameters estimated from the fMRI data, but that did not have an impact on the result. Additionally, we looked at cortical thickness differences. We observed again a statistically significant effect: increased cortical thickness among subjects with ASD. Again, this was exclusively driven by the NYU data.

In conclusion, our study was not able to replicate the findings reported by Andrews et al (2017). We also observed that the NYU site in ABIDE seems to have a yet undiscovered artefact which produces statistically significant differences in grey-white contrast and cortical thickness. It could be argued that the failure to reproduce the original finding while excluding NYU was due to the heterogeneity of ABIDE. However, we were not able to reproduce the finding in the EU-AIMS cohort either, where every effort has been made to harmonise inclusion criteria and MRI sequences. Until the cause for this difference is discovered, the NYU data should be treated with special attention. In order to facilitate tracking this artefact down, we have made our code available on GitHub: https://github.com/neuroanatomy/GWPC.

The shape of primate brains varies widely from the small lissencephalic lemurs to the large and richly folded brains of great apes. Studying this morphological diversity across phylogeny allows us to better understand the way in which primate brains adapt, and in particular, the evolutionary context of the human brain. Recent advances in generative machine learning models have led to algorithms capable of learning shape embeddings and to generate realistic new instances (Park et al. 2019). We explored the utilisation of an autoencoder deep neural network to generate shapes of primate brains, based on a sample of close to 100 individual brains from 34 different species from Heuer et al. (2019).

Our network successfully learnt a landscape of changes in shape. Interestingly, species with brains of comparable volume were close in the learnt space, despite having been size-normalised for the training. This suggests that changes in volume are consistently concomitant with changes in shape, and this disregarding the species’ position in the phylogenetic tree. Indeed, brains with large volumes tend to be rounder. This seems to be due to a more pronounced dorsal development in large brains, the orbital region occupying progressively a more ventral position.
Our network is also able to generate new data. We generated possible brain shapes for all ancestral states of the primate phylogenetic tree, and evolutionary trajectories for each of our species back to the brain of the common ancestor.

In conclusion, we successfully trained a deep neural network to learn the space of morphological variation across primates, and to generate new data along the phylogenetic tree. We obtained evolutionary trajectories of extant primate brains all the way back to the common ancestor.

The development of neocortical folding is concomitant with that of important developmental landmarks such as cytoarchitectonic regionalisation and cortico-cortical connectivity. Brain folding is likely to result from a buckling instability produced by the rapid growth of the neocortex on top of the subcortical substrate (the developing white matter). Brain folding does not only lead to the emergence of gyri, but also forms surprisingly stable folding patterns. While the mechanical nature of brain folding has been increasingly accepted, the stability of the folding pattern is still believed to be mostly genetically encoded.

We investigated the ability of a biomechanical model to produce a stable folding pattern by simulating the folding of the ferret brain. The initial geometry of our model was obtained from newborn ferret MRI and histological data. We modelled the neocortex, white matter and other subcortical regions as hyper-elastic materials, with Young modulus and Poisson ratios similar to those observed in real developing brains. The growth of the neocortex and subcortical regions had a component independent of stress (morphogenetic growth), of a rate similar to that of real ferrets. Morphogenetic growth was homogeneous. Additionally, the neocortex, white matter and subcortex were also plastic, i.e. they grew or contracted in response to stress.

We observed the formation of brain folding with a pattern that reproduced some of the major traits of the real ferret folding pattern. The ferret brain is longer than wide, and folds tend to follow this axis. The main influence on folding wavelength was neocortical thickness, which resulted in a number of folds similar to that observed in real ferrets. Overall, most of the characteristic sulci of the ferret brain could find their counterpart in our simulations.

Mechanical forces have been shown to have strong and varied effects on the developing tissue, including the ability to influence cell proliferation, apoptosis, cell fate, cell shape and connectivity. The mechanical forces induced by brain folding could then have a causal effect on the organisation of the neocortex. We have shown that because of constraints due to the initial neocortical geometry, also the pattern of these folds could be mechanically determined. This process could be a main mechanism through which development is able to produce a robust neocortical organisation, and a way in which an organ-wide layout of stereotypical cytoarchitectonic and connective modules could be added or removed through evolution.

Understanding others is fundamental to interpersonal coordination and successful cooperation. One mechanism posited to underlie both effective communication and behavioral coordination is interpersonal neural synchrony. Although presumably foundational for children's social development, research on neural synchrony in naturalistic caregiver-child interactions is lacking. Using dual-functional near-infrared spectroscopy (fNIRS), we examined the effects of interaction quality on neural synchrony during a problem-solving task in 42 dyads of mothers and their preschool children. In a cooperation condition, mothers and children were instructed to solve a tangram puzzle together. In an individual condition, mothers and children performed the same task alone with an opaque screen between them. Wavelet transform coherence (WTC) was used to assess the cross-correlation between the two fNIRS time series. Results revealed increased neural synchrony in bilateral prefrontal cortex and temporo-parietal areas during cooperative as compared to individual problem solving. Higher neural synchrony during cooperation correlated with higher behavioral reciprocity and neural synchrony predicted the dyad's problem-solving success beyond reciprocal behavior between mothers and children. State-like factors, such as maternal stress and child agency during the task, played a bigger role for neural synchronization than trait-like factors, such as child temperament. Our results emphasize neural synchrony as a biomarker for mother-child interaction quality. These findings further highlight the role of state-like factors in interpersonal synchronization processes linked to successful coordination with others and in the long-term might improve the understanding of others.

Musical improvisation is a sophisticated cognitive process that combines creativity, goal-directed action, sensory monitoring, and social interaction. With a renewed interest in quantifying creative processes facilitated by recent advances in neuroimaging technology, musical improvisation has emerged as an ideal paradigm to study creativity. However, many studies isolate the top-down processes related to creativity from those related to production and auditory perception, leaving the question of how creative behaviours integrate sensory information with higher cognitive processes unanswered. The bottom-up neural correlates of music perception have been extensively quantified, comprising networks for auditory processing and parsing semantic and syntactic content. In studies of spontaneously-generated music and domain-general creativity, executive control and goal-directed movement networks are added to the perceptual foundation. This presentation will summarize previous work on music perception and improvisation, and will present a conceptual model of musical improvisation with known neural correlates. We make recommendations on future directions for the study of improvisation and discuss the challenges posed by this endeavour.

The use of advanced neuroimaging analysis methods (i.e. graph theory, machine learning) in neuroimaging has exponentially increased in recent years and many of these techniques have been applied to clinical neuroimaging datasets. However, as techniques have become more complex, communication between clinicians and researchers has become more challenging and thus translational research could be weakened. In my observation many questions formulated in studies are not relevant in clinical care and also many questions from clinical care could be addressed more in clinical neuroimaging research. In my twitter talk, I will therefore talk about how to formulate clinically relevant research questions in neuroimaging by reviewing successful examples from a point of view of a clinican. I will introduce the PICO technique from evidence-based medicine, which can be used to systematically create questions for clinical medicine. I show through examples how this system can be applied to clinical neuroimaging studies. This talk will give researchers the opportunity to better review their own research questions in clinical neuroimaging and help them to find new, relevant questions through the perspective of clinical research.

Introduction

Controlling craving is key to quit smoking. To help smokers tolerate nicotine craving better, we trained them with neurofeedback to downregulate cue-induced craving activation in the anterior cingulate cortex (ACC). Neurofeedback training was accomplished using a novel adaptive real-time fMRI neurofeedback paradigm where the intensity of the displayed nicotine cues was dynamically coupled to ongoing ACC activity.

Materials and Methods

64 nicotine dependent subjects who wanted to reduce their weekly cigarette consumption or quit smoking completely participated in the study. They were randomly assigned to either the experimental group (EG, N=32) or a control group (CG, N=32).

Subjects in the EG were trained to downregulate their ACC activity, an area that responds to craving-related nicotine cues. In the CG, feedback was linked to activity in the angular gyrus, which is not associated with nicotine craving. During the neurofeedback runs, subjects were presented stimuli whose craving intensity was coupled to brain activity in either the ACC (EG) or the angular gyrus (CG). The better subjects were at down-regulating the targeted brain activity, the more intense were the presented nicotine cues (i.e., the task became more difficult in the EG).

Subjects were trained for 10 neurofeedback runs (4.5 mins each), spread over two neurofeedback sessions within one week. Before and after neurofeedback training, clinical and behavioral assessments were performed, including a follow-up session five weeks after the last neurofeedback session.

Results and Conclusions

For subjects in the EG, the average number of weekly consumed cigarettes decreased from 83.02 to 43.27 (p<0.001). 75% of the EG subjects reduced cigarette consumption, with 18.75% of them quitting smoking completely. Fagerström dependence scores were reduced significantly (p<0.001). Finally, craving ratings of nicotine cues decreased significantly and correlated with decreased cigarette consumption (p < 0.01). In the CG, weekly cigarette consumption decreased from 84 to 68.93 (p=0.084). Here, 60% of the subjects reduced cigarette consumption and 10% quit smoking completely. Fagerström dependence scores were decreased (p = 0.090). Overall, cigarette consumption and Fagerström dependence scores were significantly more reduced in the EG compared to the CG (cigarettes: p=0.029, Fagerström: p=0.005). Currently ongoing analyses of pre- and post-training resting-state and fMRI-craving-task data as well as the neurofeedback training and transfer run data will help better understanding the neuroscientific underpinnings of these large clinical effects.

Adaptive neurofeedback training supports smoking cessation by reducing craving when being confronted with nicotine cues. Our results suggest that brain-controlled adaptive nicotine cue exposure stimulation might be promising novel therapeutic tool in addiction.

Real-time fMRI-based neurofeedback is an emerging scientific as well as clinical tool that allows for learning to self-regulate brain activity. It has been shown to modulate behavior in healthy individuals, and further has demonstrated the capacity to improve clinical symptoms in various patient populations. However, the performance in neural activity self-regulation varies considerably across studies and individuals. Consistent learning curve patterns, such as steadily rising regulation performances across runs, are rare. Here, we investigate whether neurofeedback regulation performances across runs of studies are merely random or follow a predictable pattern. This is achieved by applying machine-learning (L1-regularized Linear Regression and ensemble of Randomized Trees) to predict the regulation performance of a training run based on previous training run performances. In our models, we included subject- and study-specific characteristics such as age, sex, instructions, trained brain regions, or the length of regulation blocks to investigate whether these factors affect performance. Permutation-based importance rankings were computed for each model input. Preliminary evidence in a sample of 170 participants from 11 studies (10 participants were excluded as they were identified as being outliers by Isolation Forests), suggests that we are able to predict regulation performance to some degree (Lasso: median R² up to 0.18; Randomized Trees: median R² up to 0.1). Regarding the feature importance analysis both the Lasso and the Randomized Trees show the same pattern. The previous run performances are the most important predictors, whereas the most recent run is the most important predictor in 3 out of 4 cases for both models. Hence, it can be summarized that neurofeedback training is not entirely random but follows to some degree predictable patterns.

Traditional neuroimage analysis pipelines involve computationally intensive, time-consuming optimization steps, and thus, do not scale well to large cohort studies. With FastSurfer we propose a fast and extensively validated deep-learning based alternative to FreeSurfer for the automated processing of structural human MRI brain scans, including volumetric segmentation, surface reconstruction of cortical geometry, morphometric estimation of brain structures and cortical thickness. It is the first work integrating a novel deep learning method for image segmentation into a complete processing pipeline. FastSurfer consists of an advanced deep learning architecture (FastSurferCNN) used to segment a whole brain MRI into 95 classes in under 1 min, and a surface pipeline building upon this high-quality brain segmentation. For sustainability of the pipeline we perform extensive validations:

(i) Accuracy: FastSurferCNN consistently outperforms state of the art deep learning architectures across five independent test sets by a margin: final Dice Similarity Coefficient (DSC) of 89.08 and 85.88 for subcortical and cortical structures (highly significant Wilcoxon signed-rank test after Bonferroni correction).

(ii) Generalizability: FastSurfer achieves high DSC for both, subcortical and cortical structures, on unseen downsampled and defaced images (Human Connectome Project: 0.87/0.85), across different vendors (GE: 0.90/0.86, Philips: 0.89/0.86, Siemens: 0.90/0.87) and disease states (cognitive normal (CN): 0.90/0.87, Alzheimer's Disease (AD): 0.89/0.86) based on 180 cases from ADNI.

(iii) Reliability: FastSurfer exhibits high test-retest reliability, with an average Intraclass correlation coefficient of 0.9 on cortical and 0.99 on subcortical regions for 20 Test-Retest subjects from OASIS1.

(iv) Sensitivity: in a group comparison between 273 non-demented and 97 demented subjects from OASIS1 FastSurfer shows increased sensitivity relative to FreeSurfer and robustly detects reduced cortical thickness in regions associated with dementia.

(iv) Speed: a significant speed-up compared to FreeSurfer is achieved by omitting several steps that have become obsolete, such as skull stripping and non-linear atlas registration. Whole brain segmentation is available within 1 min and surface-based thickness analysis within 1h (plus optionally 30 min for group registration) compared to 4h with FreeSurfer (in parallel mode, 4 threads).

Overall we introduce a fast, stable, reliable and sensitive pipeline for automated neuroimage analysis that scales well to large datasets and enables various new applications where segmentation speed is essential, e.g. to localize structures during image acquisition, to provide quantitative measures in clinical workflows, or to process large cohort studies efficiently.

The incredible success of DL tools for processing natural images in computer vision infatuates extending their use to learn patterns in brain imaging data for diagnostic classification, regression, disease characterization and prediction tasks. DL applications in the brain imaging sphere have just about started accomplishing success; however, alike other influential technologies, there a considerable amount of hype built around them, largely in terms of their viability to learn subtle characteristics of such data at all hierarchies and unlimited scalability potential. Perhaps because of this, DL methods appear to be constantly torpedoed with critiques highlighting limitations on its viability, scalability and interpretative power, however, often by basing conclusions on performance comparisons for ill-contrived feature spaces and by very frequently avoiding emphasis of some of the key reasons DL is so useful.

Coherently probing a ten-way age and gender-based classification task on a large dataset of 12,314 structural MRI (gray matter) images, this work demonstrates that DL methods, if adopted for the appropriate endeavors and trained adeptly, have the potential to substantially outperform standard machine learning (SML) methods and to scale very well presenting a lower asymptotic complexity in relative computational time. We also show that the performance improvement (with an increase in training sample size) for DL methods plateaued similar to SML methods though with significantly higher performance. There are however many ways to potentially improve further by testing even deeper models, finetuning the existing DL models, and exploring newer DL approaches. We also demonstrate that superior feature extraction via a rigorous training procedure contributed to the superior performance of DL approaches and that SML methods can perform equally well if using the feature spaces as predictive as those encoded by DL frameworks. Finally, we highlight that DL embeddings span a comprehensible projection spectrum. and show that DL consistently localizes discriminative brain biomarkers, providing an example of the robustness of DL relevance estimates.

In sum, our findings highlight the presence of non-linearities in the brain imaging data that DL models can exploit to generate more predictive encodings for characterizing the human brain. Results are in support of the potential of DL applications to brain imaging data, even with currently available data sizes. Previous claims of the unlimited scalability of DL approaches, however, demands further confirmation. We posit that future DL work in brain imaging should focus on exploring superior predictive encodings and enabling more precise discriminative feature localization through methodical model interpretations. Rather than focusing on ways to show DL does not predict as well in certain cases, we should be leveraging the flexibility of these models, to solve problems that SML approaches are not able to do.

Functional connectivity (FC) from task and rest fMRI can jointly improve
the prediction of fluid intelligence at the single subject level, so we
test if this holds for other phenotypes. Here, we used rest-FC and
task-FC to predict behavioral measures across three behavioral domains
(cognition, personality, mental health) in the Adolescent Brain
Cognitive Development (ABCD) study. In doing so, we establish the
existence of shared and unique network features in children that support
prediction across these behavioral domains.

We used minimally
preprocessed rest-fMRI and task-fMRI data from the ABCD 2.0.1 release.
We examined 3 fMRI tasks: monetary incentive delay (MID), N-Back and
stop signal task (SST). After quality control, there were 1858 unrelated
healthy children with data from all fMRI states (rest, MID, N-Back,
SST) and 36 cognitive, personality and mental health measures. We used
400 cortical ROIs from the Schaefer parcellation and 19 subcortical ROIs
from the Freesurfer Aseg atlas to generate 419 x 419 FC (Pearson’s
correlation) matrices for each fMRI state.

We used kernel ridge
regression (KRR) to predict each behavior. The input to KRR is a NxN
similarity (kernel) matrix, where N is the number of subjects. KRR
assumes that the behavior of a test subject is more similar to the
behavior of a training subject if their brain measurements are more
similar. Inter-subject similarity was computed by correlating subjects’
FC matrices, yielding one NxN kernel matrix for each fMRI state. We used
each kernel matrix on its own for single KRR. The 4 kernel matrices
were also jointly used for prediction using multi-kernel regression. KRR
was done with a nested cross-validation procedure with 120 different
data splits. Participants were divided into 10 sets. Children from the
same site were kept together. Within each split, we trained the KRR and
tuned its hyperparameters on 7 random sets, then applied KRR to the 3
remaining sets.

Of the single kernel models, N-Back performed
the best for all behaviors, but was only significantly better than
rest-FC for cognition, but not other domains. Combining task and rest
(multi-kernel FC) improved prediction for cognition and personality.
After permutation testing, multi-kernel FC predicted 32 behavioral
measures better than chance (mean cross-validated accuracies, measured
as correlations between predicted and true values, ranged from 0.05 to
0.53).

To interpret the involvement of each FC edge in the
prediction of each behavior for each fMRI state, we computed 419 x 419
predictive-feature matrices. The matrices were highly similar across
fMRI states. The matrices were also more similar within each behavioral
domain than across behavioral domains, suggesting a shared set of
network features driving the prediction of measures within each
behavioral domain. Cognition, personality and mental health were
supported by distinct features among connections from the default,
dorsal attention, and control networks.

Neuroethology is the study of brain and behavior in the context of an animal’s natural environment [1]. In behavioral simulations, this naturalism can be approximated through interdependencies among brain, body, and environment. Our approach is particularly interesting to brain scientists because simulated environments allow for environmental and neurological components of a naturalistic interaction to be both specified and controlled. We use a type of Braitenberg Vehicle [2] called Developmental Braitenberg Vehicles (dBVs) to model the developmental aspects of an embodied artificial nervous system. As models of developmental origins, dBVs extend the understanding and functional context of original Braitenberg Vehicle behaviors. These behaviors range from phototropisms and sensorimotor coordination to more complex behaviors that resemble emotional valence and acquired knowledge. As the preview of a forthcoming paper from our group on this topic [3], we propose three distinct approaches to dBVs in software. These include a model based on genetic algorithms, a multisensory Hebbian learning model, and a multi-agent (dBV collective) model. Our software approaches can be used to study many different behavioral scenarios. Three instances of these involve optimized spatial cognition (genetic algorithm model), hinges connecting behavioral and neural models (multisensory Hebbian learning model), and cumulative classification (multi-agent approaches).


[1] Datta, S.R., Anderson, D.J., Branson, K., Perona, P., and Leifer, A. (2019). Computational Neuroethology: a call to action. Neuron, 104(1), P11-P24.

[2] Braitenberg, V. (1984). Vehicles: experiments in synthetic Psychology. MIT Press, Cambridge, MA.

[3] Dvoretskii, S., Gong, Z., Gupta, A., Parent, J., and Alicea, B. (2020). Braitenberg Vehicles as Developmental Neurosimulation. In progress.

Introduction: There is recent debate on the use of gradients versus parcellations for representing resting-state functional connectivity (RSFC) data. There is also significant interest in using RSFC to predict behavior. Here, we compare 2 gradient techniques (Gordon2016; Margulies2016), 2 soft-parcellation techniques (Beckmann2004; Harrison2019) and 2 hard-parcellation techniques (Schaefer2018; Kong2020) for predicting 58 behavioral measures in healthy adults.

Methods: We used ICA-FIX rs-fMRI data from the HCP S1200 release. 58 cognitive, personality and emotion measures were selected. Subjects with 4 runs and no missing behavioral measures were considered (N=746). Kernel ridge regression (KRR) was used to predict each behavior. The input to KRR is an "N×N" similarity (kernel) matrix, where N is the number of subjects. KRR relies on the idea that the behavior of a test subject is more similar to the behavior of a training subject if their brains are more similar. For Gordon2016, inter-subject similarity was defined as the correlation between subjects’ gradient maps. Thus, each approach yields a single "N×N" similarity matrix for prediction. For Margulies2016, inter-subject similarity was defined as the correlation between 5 gradient maps, yielding 5 "N×N" similarity matrices, which were used by multi-kernel regression for prediction. For Harrison2019, we defined two similarity matrices. The first is the correlation between the subject’s partial correlation matrices. The second was defined as the correlation between each of 50 spatial maps, yielding 50 similarity matrices that were averaged, yielding a single "N×N" similarity matrix. The two "N×N" similarity matrices were used by multi-kernel regression for prediction. For Beckman2004, Schaefer2018 and Kong2020, inter-subject similarity was defined as the correlation between subjects’ partial correlation matrices. We performed 20-fold cross-validation: KRR was trained on 19 folds and used to predict behavior in the test fold. Regularization parameters were determined using inner-loop cross-validation. The 20-fold cross-validation was repeated 100 times to ensure stability.

Results: By comparing prediction accuracies averaged across 58 behavioral measures and 100 cross-validation splits across approaches, we found that soft-parcellations were as good as or better than gradient techniques, hard-parcellations were as good as or better than soft-parcellations. Similar conclusions were obtained with coefficients of determination (COD).

Conclusions: In a kernel regression framework, RSFC derived from brain parcellations were better than gradient techniques for behavioral prediction. However, we caution that gradient techniques are relatively new, so future research might yield further improvements. Furthermore, Kong2020 actually incorporated gradients in deriving the individual-specific parcellations. Experiments with more datasets and other supervised learners would also be desirable.

Introduction
Mental disorders such as major depressive disorder (MDD) are characterized by imbalances in homeostatic processes reflected in altered neural function at rest. According to the Embodied Predictive Interoceptive Coding model (EPIC), agranular regions generate predictions of upcoming stimuli based on previous experience while granular regions compute prediction errors based on incoming sensations. Within the insular cortex, distinct subregions have been associated with predictive or evaluative coding of homeostatic states, respectively. However, it is still elusive if functional segregation within insular subregions is impaired in MDD patients. Based on the EPIC model, we expected a loss of specificity in insular functional connectivity (FC) in MDD which reflects an increased focus on internally generated predictions and a reduced attention to sensory input.

Methods
To test for loss of specificity within FC profiles of insular subregions in MDD, we measured resting-state fMRI in 433 MDD patients and 417 healthy control (HC) participants. Using the Hammers atlas, we calculated resting-state FC (seed to voxel) from six (three anterior and three posterior) subregions. We tested for specificity by comparing the similarity across subregions and within vs. between anterior and posterior insular FC profiles for MDD and HC using Pearson correlation. To ensure that our results are spatially specific, we performed the same calculations in the temporal cortex as a reference region. To examine group-specific shifts in insular subregional FC which might underlie differences in specificity, we further applied the shift function. This method compares quantiles of FC distributions of participant groups for each insula subregion.

Results
As expected, MDD patients showed a greater similarity in FC profiles of insular subregions compared to HC (t = -4.31; p < .001) suggesting lower functional segregation. Critically, this difference in similarity was specific to the insular cortex as comparing FC profiles of 6 subregions within the temporal cortex yielded no significant interaction effect (t = -1.37; p = .17). This insula-specific loss of functional segregation was driven by an increase in similarity within subregions of the anterior insula compared to the posterior insula (t = 4.09, p < .001). Crucially, mean FC for anterior insular subregions to the whole brain was shifted towards hyperconnectivity in MDD.

Conclusions
Congruent with theories of aberrant interplay between interoceptive predictions and sensory signals, our results corroborate the hypothesized loss of functional segregation within the insula. We conclude that an increased reliance on interoceptive predictions and a discounting of exteroceptive signals might contribute to the pathophysiology in MDD.

INTRODUCTION. Epilepsy is increasingly conceptualized as a network disorder. Here, we integrated neuroimaging and connectome analysis to identify network mechanisms underlying atrophy patterns in 1,021 epileptic patients relative to 1,564 healthy controls from 19 international sites.

METHODS. We studied two patient subcohorts with site-matched healthy controls: temporal lobe epilepsy with hippocampal sclerosis (TLE; nHC/TLE=1,418/732) and idiopathic/genetic generalized epilepsy (GE; nHC/GE=1,075/289). All participants underwent a T1w scan and had cortical thickness and subcortical volume measures for 68 cortical and 14 subcortical regions. Data were harmonized across 19 international sites, and statistically corrected for age, sex, and intracranial volume.

Using linear models, we first established patterns of atrophy in TLE and GE. Next, we evaluated whether these abnormalities were guided by normative network organization. To this end, we derived functional and structural connectivity matrices from the HCP dataset and computed cortical and subcortical degree centrality. Brain regions that were most vulnerable to structural compromise, as well as those considered disease epicenters, were identified via spatial correlations between atrophy maps and structural and functional (i) degree centrality and (ii) region-specific connectivity profiles. Taking advantage of large cohorts with heterogenous duration of epilepsy and a wide age range, we then separated out epilepsy-related progression from normal aging and evaluated the influence of connectome organization disease progression. Lastly, we adapted our network-based models to assess whether connectivity organization shapes atrophy patterns in individual patients.

RESULTS. Hub regions were more susceptible to epilepsy-related atrophy in both TLE and GE, with cortical hubs being most affected in the former and cortico-subcortical hubs being most vulnerable in the latter. Using network information to decode atrophy patterns, we localized disease epicenters in mesiotemporal and limbic cortices in TLE as well as sensorimotor cortices and subcortical areas in GE. Critically, these findings support diverging influences of network architecture on epilepsy-related atrophy in focal and generalized epilepsies, with subnetwork effects aligning with the pathophysiology of each syndrome. Assessing markers of disease progression revealed a similar network-level dichotomy between TLE and GE, indicating stronger influence of connectome architecture on the disease unfolds in TLE. Moving beyond group-level atrophy patterns, we used a patient-tailored adaptation of our network-based models and confirmed that relationships between atrophy and normative connectivity organization were translatable to individual patients.

CONCLUSIONS. Through worldwide collaboration in ENIGMA Epilepsy, we offer a novel network-based mechanistic perspective that can help to understand the pathological cascades in the common epilepsies.

Schizophrenia has been conceived as a disorder of brain dysconnectivity, however, the biological mechanisms underlying this network phenotype are still poorly understood, and progress in new therapeutics has been correspondingly limited. Recently, we proposed a new approach called 'morphometric similarity' to study brain structure, which characterises the similarity between brain regions in terms of their morphology (Seidlitz et al, 2018), measured using MRI. Morphometric similarity has already been shown to capture the underlying cytoarchitecture of the brain, and correlated with connectivity derived directly from tract tracing in the macaque monkey (Seidlitz et al, 2018). Here we investigated how morphometric similarity differed between the brains of patients with psychotic disorders and healthy control subjects.

We used three independent case–control MRI studies of psychosis: in total, n = 185 cases with psychotic disorders and n = 227 healthy control subjects (Habets et al, 2011, Cetin et al, 2014). For each subject, we used Freesurfer to extract regional estimates of 7 metrics: gray matter volume, surface area, cortical thickness, Gaussian curvature, mean curvature, FA, and mean diffusivity; in each of 308 cortical regions. We then correlated these morphometric feature vectors pairwise between cortical regions, to form a 308×308 morphometric similarity matrix Mi for each participant, i=1,…N.

Psychosis was associated with globally reduced morphometric similarity in all three studies, meaning that patients' brain regions were more differentiated from each other than observed in control subjects. This global decrease emerged from a broadly replicable cortical pattern of regional differences in morphometric similarity, which was significantly reduced in patients in frontal and temporal cortical areas but increased in parietal cortex.

In the second part of the study, we related this replicable cortical map of case-control differences to prior brain-wide gene expression data from the Allen Human Brain Atlas (Hawrylycz et al 2012), in an effort to connect these MRI phenotypes to the emerging genetics and functional genomics of schizophrenia. The cortical map of case–control differences in morphometric similarity was spatially correlated with cortical expression of a weighted combination of genes enriched for neurobiologically relevant ontology terms and pathways. In addition, genes that were normally overexpressed in cortical areas with reduced morphometric similarity were significantly up-regulated in three prior post mortem studies of schizophrenia.

We propose that this combined analysis of neuroimaging and transcriptional data provides insight into how previously implicated genes and proteins as well as a number of unreported genes in their topological vicinity on the protein interaction network may drive structural brain network changes mediating the genetic risk of schizophrenia.

Background and aims: Alzheimer's disease is a multifactorial dementia disorder characterized by early amyloid-β, tau deposition, glial activation and neurodegeneration. There are differences between Familial and Sporadic AD in terms of age of onset as well as cognitive profile and patients’ clinical presentation. In this study, we aimed to analyze difference of amyloid-β burden’s brain locations between Familial and Sporadic Alzheimer’s patients.
Methods: Clinical and imaging data of 17 familial and 18 sporadic Alzheimer's patients, that have no difference between the two groups in terms of sociodemographic characteristics and disease duration, from Alzheimer's Disease Neuroimaging Initiative(ADNI) were included in the study. Early-onset FAD (EOFAD),17 patients, based on the underlying genetic mechanism are: 4 patients have mutation of APP, 11 patients have mutation of PSEN1, 2 patients have mutation of PSEN2. Flutemetamol radionuclide(F-18) marking amyloid PET images were used and analyzed with VINCI (“Volume Imaging in Neurological Research, Co-Registration and ROIs included”).
Results: There was no significant difference between the two groups in terms of total amyloid burden. In the familial group, the amyloid burden was higher in the insular cortex, striatum, supra marginal gyrus, orbitofrontal cortex and cingulate cortex than in the sporadic group.
Conclusion: The findings of our study support other studies suggesting that frontal and extrapyramidal amyloid burden is higher in familial AD cases compared to sporadic cases.

Dorsal and ventral premotor cortical areas play different roles in motor control. Their structural connections have been tested in the non-human primate, but not yet in humans. We quantified, for the first time, the effect of direct electrical stimulation (DES) applied on the subcortical level, producing transient disruption to ongoing object manipulation in the awake neurosurgical procedures of thirty-six brain tumour patients. DES caused anatomically dissociable effects on muscle activity patterns: movement arrest occurred in dorsal premotor white matter and ‘clumsy’ movement was produced in ventral premotor white matter. Using diffusion tractography to examine the perturbed structural connectivity, we show that movement arrest is associated primarily with connections of the superior frontal gyrus, while discoordination was associated with the ventral fronto-parietal tract (SLF III). These results causally demonstrate the involvement of premotor pathways in the orchestration of hand muscles required for motor control, likely mediating hierarchical stages of motor processing.

We propose a method to characterise network dynamics in MEG resting-state data, combining a pairwise maximum entropy model (pMEM) and the associated energy landscape analysis. We apply this method to patients suffering from juvenile myoclonic epilepsy (JME) and healthy control group. Fifty-two subjects in total underwent a resting-state MEG recording session. We fitted the pMEM to the oscillatory power envelopes in theta (4-7 Hz), alpha (8-13 Hz), beta (15-25 Hz) and gamma (30-60 Hz) bands in 3 source-localised resting-state networks: the frontoparietal network (FPN), the default mode network (DMN), and the sensorimotor network (SMN). The pMEM provided an accurate fit to the MEG oscillatory activity in both patient and control groups, and allowed estimation of the occurrence probability of each network state, with its regional activity and pairwise regional co-activation constrained by empirical data. We used energy values derived from the pMEM to depict an energy landscape of each network, with a higher energy state corresponding to a lower occurrence probability. When comparing the energy landscapes between groups, JME patients showed fewer local energy minima than controls and had elevated energy values for the FPN within the theta, beta and gamma-bands. Our findings suggested that JME patients had altered multi-stability in selective functional networks and frequency bands in the frontoparietal cortices.

While there is an overall decline in cognitive functioning with age, some individuals maintain similar cognitive performance to that of young adults, while others become impaired. I will discuss how small differences in white matter (WM) integrity influence individual differences in the maintenance and updating of both resting-state and task-related networks over time. EEG was recorded while young (20-35 years) and healthy older participants (60-75 years) performed a task specifically designed to engage working memory and inhibitory processes, and the association between functional activity. We also assessed the degree to which an older individual’s tract microstructure differs from the microstructural profile of young adults with DTI. Source reconstruction was cortically constrained to anatomical data and phase-locking value was used to assess the functional synchrony between brain regions. Graph theory analyses were applied to each frequency band over sliding time windows to study age effects on dynamic frequency-specific networks and their association with the integrity of the underlying structural network. We also relied on the CamCAN (Cambridge Centre for Ageing and Neuroscience) dataset to combine DTI and MEG data and investigate the spatial and oscillatory characteristics of resting-state brain networks, and to replicate our results in another independent dataset. With age, alterations of the structural network were associated with a reduced maintenance of the modular organization of functional networks, delayed task-related effects, and larger variability across time windows. These effects were mainly observed in the alpha and gamma bands. Altered network modularity was also associated with declined working memory performance during aging. By applying network analyses to MEG/EEG source and DTI tract-specific integrity, new insights can be obtained on how the variability of white matter alterations underlies age-related changes of functional network integrity and cognitive performance. These results improve our understanding of the mechanism by which age-related white matter alterations impact functional network integrity and subsequently results in poorer cognitive performance.

Obtaining accurately segmented white matter bundles in the human brain is essential for multiple applications, such as surgical planning in neurosurgery and group studies in cognitive neuroscience. Having a good and reliable segmentation is however not trivial, mainly because of the intrinsic complexity of the data. In the last decade, multiple methods have been developed to automatically segment bundles from diffusion Magnetic Resonance Imaging (dMRI) data. Although notable improvements have occurred over the years, the quality of segmentation is not yet satisfactory. Most of the current methods require extensive neuroanatomical knowledge, are time consuming, or are not robust to different data settings, such as when there are differences in bundle size, tracking algorithm, and/or quality of dMRI data.

To overcome these limitations, we propose a novel supervised streamline-based segmentation method, called Classifyber, which combines in a linear model both geometrical and anatomical features of the streamlines. In particular, Classifyber implements a linear classifier that accurately predicts whether or not a given streamline belongs to the bundle of interest, by leveraging multiple example bundles segmented by experts. To this end, it combines both the similarity measures between streamlines, typical of streamline-based methods, and the anatomical information from Regions of Interest (ROIs), typical of ROI-based and connectivity-based methods.

We compared the proposed Classifyber against three state-of-the-art automatic bundle segmentation methods: TractSeg, RecoBundles and LAP. Four different tractography datasets were employed, which differed in terms of bundle size (large vs. small), tracking algorithm (probabilistic vs. deterministic), and dMRI data quality (research vs. clinical). Ground truth bundles were either manually segmented by experts or by using semi-automatic techniques followed by visual inspection. Across the segmentations of 18 different kinds of bundles in hundreds of subjects, Classifyber obtained the highest Dice Similarity Coefficient (DSC) scores by a substantial margin, which on average was 0.85±0.05, followed by LAP (0.75±0.06), TractSeg (0.71±0.13) and RecoBundles (0.64±0.14).

Classifyber provides evidence to substantially improve the quality of segmentation with respect to state-of-the-art methods and, more importantly, it is robust to different data settings, i.e. across different kinds of bundles, tractography techniques, expert-made segmentations, and quality of dMRI data. Moreover, the proposed method is fast to compute and it is freely available as an open source web application through the platform BrainLife. When using the pre-trained algorithm, the segmented bundle of interest can be obtained in a few minutes, thus providing a powerful segmentation tool that can be appreciated also in clinical settings.

Sophisticated visualization tools are essential for presentation and exploration of human neuroimaging data. While two-dimensional orthogonal views of neuroimaging data are conventionally used to display activity and statistical analysis, three-dimensional (3D) representation are useful for showing the spatial distribution of a functional network, as well as its temporal evolution. For these purposes, there is currently no open-source, 3D neuroimaging tool that can simultaneously visualize desired combinations of MRI, CT, EEG, MEG, fMRI, PET, and intracranial EEG (i.e., ECoG, depth electrodes, and DBS). Here we present the Multi-Modal Visualization Tool (MMVT), which is designed for researchers wishing to better understand their neuroimaging functional and anatomical data through simultaneous visualization of these existing imaging modalities. MMVT contains two separate modules: The first is an add-¬on to the open-source, 3D-rendering program Blender. It is an interactive graphical interface that enables users to simultaneously visualize multi-modality functional and statistical data on the cortex and in subcortical surfaces as well as the activity of invasive electrodes. This tool also enables highly accurate 3D visualization of neuroanatomy, including the location of invasive electrodes relative to brain structures. The second module includes complete stand-alone pre-processing pipelines, from raw data to statistical maps. Each of the modules and module features can be integrated, separate from the tool, into existing data pipelines. This gives the tool a distinct advantage in both clinical and research domains as each has highly specialized visual and processing needs. MMVT leverages open-source software to build a comprehensive tool for data visualization and exploration.

Analyzing the temporal reconfiguration of brain functional connectivity has become a key element to study the functional interactions between neuronal populations and to characterize their role in brain function. Although most dynamic connectivity studies consider themselves to be spatiotemporally dynamic, almost all of them only focus on temporally dynamic properties of the brain. This results in overlooking important spatially dynamic features of the brain. In this work we will describe and define spatial, temporal, and spatiotemporal dynamics. We will discuss how incorporating space into dynamic functional connectivity research can broaden our scientific knowledge of brain function.

We will also highlight work investigating spatial dynamics using a spatially fluid chronnectome model. We show evidence that moment-to-moment spatial reconfiguration of brain networks occurs at the finest observable scale (voxel-level). We observe that brain networks evolve spatially over time, and they transiently merge and separate from each other. Interestingly, our analysis shows that spatial dynamics can explain the inconsistencies observed in previous spatially static studies. For instance, different spatial patterns of the default mode network reported in previous static functional connectivity studies appear at different moments of time, highlighting inconsistencies due to methods that do not account for spatial dynamics.

We can also study the dynamic properties of the brain using the hierarchical models of brain function. In this framework, we construct the brain function as a hierarchical structure based on functional homogeneity. Different levels of the hierarchy represent different spatial scales and contain different dynamic information. We show the first evidence of spatially dynamic properties within functional domains (FDs). FDs evolve spatially over time, including changes in a region's association to a given FD over time (from strong association to complete dissociation).

Incorporating space in the dynamic analysis allows us to develop a new set of dynamic metrics, which are inaccessible using previous functional connectivity methods, to study the information of brain function. Spatial dynamic studies have also revealed new patterns of alterations in patients with schizophrenia, which are hidden and undetectable using previous static and dynamic functional connectivity techniques.

Population imaging markedly increased the size of functional-imaging datasets. Analyzing these large data entails new scalability challenges, computational and statistical. For this reason, brain images are typically summarized in a few signals, for instance reducing voxel-level measures with brain atlases or functional modes. A good choice of the corresponding brain networks is important, as most data analyses start from these reduced signals. We contribute finely-resolved atlases of functional modes, comprising from 64 to 1024
networks. These dictionaries of functional modes (DiFuMo https://parietal-inria.github.io/DiFuMo/) are trained on millions of fMRI functional brain volumes of total size 2.4TB, spanned over 27 studies and many research groups. We demonstrate the benefits of extracting reduced signals on our fine-grain atlases for many classic functional data analysis pipelines: stimuli decoding, standard GLM analysis , resting-state functional-connectomes biomarkers, data compression and meta-analysis. In each of these analysis scenarii, we compare the performance of our functional atlases with that of other popular references, and to voxel-level analysis. Results highlight the importance of using high-dimensional “soft” functional atlases, to analyse brain activity while capturing its functional gradients. With these, analyses achieve similar statistical performance as at the voxel level, but with much reduced computational cost and higher interpretability. In addition, we provide meaningful names for these modes, based on their anatomical location.

Introduction
Understanding how higher order cognitive function emerges from brain connectivity depends on quantifying how the behavior of brain regions is integrated within the broader cortical landscape. Recent conceptual and methodological advances have provided both the data and analytic approaches that allow high dimensional neuroimaging and connectivity data to be represented in lower dimensional manifolds, also known as cortical “gradients” [1].
The growth in our capacity to map cortical gradients, coupled with their promise of a better understanding of how structure gives rise to function, highlights the need for a set of tools that support the analysis of neural manifolds in a compact and reproducible manner. Here, we present BrainSpace [2], an open-access set of Python/Matlab tools that allow the identification, visualization, and analysis of macroscale gradients of brain organization.

Methods
The core functionality of BrainSpace is to compute a lower dimensional manifold from a high dimensional connectivity matrix. In short, for an input matrix with n seeds (rows) and m features (columns) we compute an n-by-n affinity matrix with each element denoting the similarity of each seed. Next, we apply a dimensionality reduction technique to extract vectors, or gradients, that explain the most variance in the data. These gradients may then be used in subsequent analyses of, for example, structure-function associations.
Further functionality includes 1) inter-gradient alignments which allow for the comparison of gradients derived from separate datasets, 2) data visualizations for gradient data, and 3) null models for testing significance of associations between gradients and other markers.
BrainSpace is freely available in both Python and Matlab (https://github.com/MICA-MNI/BrainSpace), two widely used languages in the neuroimaging community, and an extensive documentation is available at https://brainspace.readthedocs.io/.

Results
We show a working example of computing functional connectivity gradients with BrainSpace, presented in both Python and Matlab. Resultant gradients can be easily plotted to a surface and used in subsequent analyses.

Conclusions
We presented a set of tools for the computation of connectome gradients. Our work is fully open-source, with all code freely available on Github. We hope that BrainSpace will make gradient analyses more accessible to the neuroimaging community, contribute to reproducible neuroscience, and enhance the quality of the analyses in this new and expanding branch of neuroimaging research.
References
[1] D. S. Margulies et al., “Situating the default-mode network along a principal gradient of macroscale cortical organization,” Proc Natl Acad Sci U A, vol. 113, no. 44, pp. 12574–12579, 2016.
[2] R. Vos de Wael et al., “BrainSpace: a toolbox for the analysis of macroscale gradients in neuroimaging and connectomics datasets,” bioRxiv, 2019.

While hippocampus proper has a unique three-layered allocortical organization, the cytoarchitectural layout of adjacent mesiotemporal regions transitions towards a six-layered architecture seen in other neocortical structures. Although a better understanding of this confluence may potentially provide important insights into cortical evolution, development, and susceptibility to disease, the field lacks computational models of the confluence of iso and allocortex.

An ultra-high resolution Merker stained 3D volumetric histological reconstruction of a post mortem human brain was obtained from the open-access BigBrain repository, along with pial and white matter surface reconstructions. We also obtained manually segmented hippocampal subregions, which were labelled with a three-way internal coordinate system; anterior-posterior, proximal-distal and inner-outer. We generated continuous inner and outer surfaces of the hippocampus by triangulating coordinates at the minimum and the maximum of an inner-outer axis. Next, we designated the most proximal coordinates as hippocampal bridgeheads (ie: closest to isocortex) and generated links to the nearest isocortex coordinates that were inferior to the bridgehead. We generated 14 equivolumetric surfaces between the inner and outer confluent surfaces and sampled the intensities from the 40um BigBrain along 9432 matched vertices, creating microstructure profiles in the direction of cortical columns for the hippocampus and adjacent mesiotemporal areas. We calculated microstructure profile covariance between all vertices and employed diffusion map embedding to determine the principle axis of cytoarchitectural variation. Spearman correlations tested the correspondence of the principle gradient with the proximal-distal axis. 20 cytoarchitectural features (depth-wise intensities and central moments) were fed into random forest regression to predict proximal-distal axis location. Feature importance was estimated as the mean decrease in variance, and features with higher than average importance were selected. The procedure was repeated across 100 splits to assess prediction robustness and feature importance.

The principle gradient of cytoarchitectural differentiation within the confluent model was closely aligned with the proximal-distal axis (r=0.76). Within hippocampus, CA2 expressed a different, a more isocortex-like cytoarchitecture than expected by its position on the proximal-distal axis. Microstructure profiles and the corresponding central moments were strong predictors of the proximal-distal axis (R2=0.91±0.004), which was driven by profile skewness, intensities at surface 3 and 9 (restricted model: r=0.90±0.03).

We introduce a new approach to harmoniously model the transition between iso- and allocortex in the mesial temporal lobe. These findings provide novel insights into the unique, and gradual cytoarchitectural properties of the human temporal lobe.

Increasing the reproducibility of neuroimaging measurement addresses a central impediment to the advancement of human neuroscience and its clinical applications. Recent efforts demonstrating variance in functional brain organization within and between individuals shows a need for improving reproducibility of functional parcellations without long scan times. We apply bootstrap aggregation, or bagging, to the problem of improving reproducibility in functional parcellation. We use two large datasets to demonstrate that compared to a standard clustering framework, bagging improves the reproducibility and test-retest reliability of both cortical and subcortical functional parcellations across a range of sites, scanners, samples, scan lengths, clustering algorithms, and clustering parameters (e.g., number of clusters, spatial constraints). With as little as six minutes of scan time, bagging creates more reproducible group and individual level parcellations than standard approaches with twice as much data. This suggests that regardless of the specific parcellation strategy employed, bagging may be a key method for improving functional parcellation and bringing functional neuroimaging-based measurement closer to clinical impact.

Functional magnetic resonance imaging(fMRI) has allowed us to detect the functional organization of the brain in detail unprecedented in the spatial resolution. Block based design tasks have predominated the study of the brain because of ability to control for confounding variables but the world we encounter is very different from the one upon which neuroscientists experiment. Since the last decade, there has been a surge of naturalistic stimuli based localization studies. In this paper, we looked at differences in the activation patterns across genders when they watched audiovisual movie stimulus(hollywood and non-hollywood) in the scanner. We have used the publicly available preprocessed Human Connectome Project high resolution 7T dataset. Out of the 184 subjects scanned, 174(107 females) took part in all four 15-min movie runs. 30 females were excluded from the dataset to match the groups. Each run had 3 clips either from Hollywood or Non Hollywood Indie movies in English language. Vertex based time series was averaged across 360 regions of interests(ROIs) or parcels, 180 per hemisphere from the HCP Multi-Modal Parcellation. The multi modal atlas also provided with a smaller set of 23 sections which comprise of the coarse parcels. This allowed for an easier neuroscientific interpretation of the results. We utilized Functional Connectivity(FC) and Inter Subject Correlation(ISC) analysis to find the presence of subtle differences across the genders in movie processing. For each value in the node x node FC correlation matrix, we performed a t-test across the groups which resulted in a node x node t-statistic map. Females were found to have a higher functional connectivity between the inferior frontal and PCC(t=-3.4,p<0.005), insula and visual regions(t=-2.9,p<0.005) and dlPFC and somatosensory/motor regions(t=-3.2,p<0.005). We used one-way ANOVA to detect FC differences for separate clips within each run and found that males show more difference in FC for individual clips in the mid cingulate and PCC/mPFC(p<0.0005) regions. Males also show higher correlated activity in the left inferior temporal(r=0.47) and right lateral frontal(r=0.48) regions. This paper demonstrates possible differences in gender based processing of naturalistic stimuli and such imaging studies should put into consideration gender during analysis of data.

The ability to reproduce published results is essential for trust in science, but full reproducibility remains challenging, requiring the sharing of code, data, and computational platform. We addressed this challenge in the Neuroimaging Analysis Replication and Prediction Study (NARPS), in which a large number of analysis teams analyzed a single fMRI dataset and submitted a broad range of results for further analysis, including thresholded and unthresholded statistical maps. The full analysis stream for these results was developed using a combination of Python and R code, in each case specifying a particular version of each software package. Analyses were implemented within a Docker container, and executed using the CircleCI continuous integration server, using data and code that are openly available for download from Zenodo, through persistent identifiers (DOIs). The results were exported as artifacts from the CircleCI server and also shared on Zenodo. This project provides a blueprint for a fully reproducible data analysis workflow, taking advantage of modern software engineering infrastructure.

Characterizing the subtle changes of MEG functional brain networks associated with the pathological cascade of Alzheimer’s disease (AD) is important for early diagnosis and prediction of disease progression prior to clinical symptoms. We developed a deep learning-based graph Gaussian embedding method, termed multiple graph Gaussian embedding model (MG2G), for identification and characterization of the early stages of AD using eye-closed resting state MEG data. Specifically, the MG2G model can learn highly informative MEG brain network node-wise patterns by encoding high-dimensional MEG resting state brain networks into a low-dimensional probabilistic space (i.e., multivariate Gaussian distributions). The MEG brain network embedding signatures enable quantitative capturing of subtle and heterogeneous node-wise brain connectivity patterns. Moreover, it can be used as input to traditional classifiers for various downstream graph analytic tasks (e.g., AD early stage prediction, statistical evaluation of between-group significant changing regions, etc.). Experimental results show that our method could provide a novel quantitative method to assess complex MEG functional connectivity patterns and extract node-wise probabilistic Gaussian distribution embeddings that are effective and convenient for AD progression prediction across different early stages of AD. Additionally, statistical evaluation results for the between-group significant brain regions (based on the Wasserstein metric) facilitate better understanding of the underlying heterogeneous pathogenesis of AD progression, and provide insights for designing precision behavioral intervention. Lastly, our method motivates a wider adoption of deep learning-based stochastic graph embedding methods to other neuroimaging data. Future work will consider extensions to multimodal data (e.g., fMRI, MEG and EEG, PET) for robust inference due to the multifaceted nature of AD.

Sudden infant death syndrome (SIDS) is the leading cause of postneonatal infant mortality in industrialized nations with a rate of 0.39/1000 live births in the United States alone. Even though some environmental factors might increase the risk of SIDS, an intrinsic defect in brain anatomy may be at the source of at least a subset of SIDS cases. Our central hypothesis posits the cause of death in SIDS as a failure in arousal in response to a life-threatening stressor. Arousal responses are mediated by the subcortical ascending arousal network (AAN). Arousal pathways of the AAN originate in the brainstem and activate awareness networks in the central cortex via synapses in the hypothalamus, thalamus, basal forebrain, or, alternatively via direct innervation of the cerebral cortex itself. Our laboratory has provided evidence over the last two decades that serotonin defects are concentrated mainly in a cytologically defined region of the rostral medullary reticular formation of the brainstem which we call the putative "core lesion of SIDS". This complex neural substrate and its interconnectivity is nearly impossible to visualize in toto with standard neuropathologic techniques at autopsy.
In this work, we show that the proposed ex vivo imaging techniques have the potential to accurately visualize the brainstem with its numerous nuclei and tracts, to localize them in the context of the full brain, as well as to reconstruct the AAN. First, we obtain structural MRI data of the whole infant brain using 7T at 200um isotropic resolution and diffusion imaging using 3T at 750um with 90 directions. Then, we block the brainstem and image it at higher resolution, 100um for the structural data on 7T and 400um and 90 directions for the diffusion data on 3T. Even at this resolution, ex vivo MRI is unable to provide sufficient contrast to delineate the brainstem nuclei and tracts due to size and lack of myelination. Thus, we rely on Optical Coherent Tomography, a 3D microscopy technique, to image the brainstem at 3.5um isotropic resolution. The acquisition of the top 100um is done prior to sectioning two 50um slices, (reserved for histology staining) and the process is repeated until the whole tissue is imaged. The intrinsic optical properties of the tissue, cells, fibers, myelin, and the high resolution provide a great contrast to identify our anatomical areas of interest. Due to low distortion, the OCT data can be used to reconstruct and segment the brainstem, to register to the high- and low-res MRI and to serve as a reference to reconstruct the histology stack. The individual segmentation of the brainstem can serve as seed regions to explore the AAN network.

Spoken word recognition requires matching a highly variable and rapidly changing acoustic stimulus to stored representations in an internal lexicon. This issue is especially important in the context of adult aging, wherein older adults report increasing difficulties relative to their younger counterparts when listening to speech. Age-related decrements in spoken word recognition on behavioral tasks have been well documented, yet the extent to which the brain networks that support spoken word recognition change with age is not currently well specified. The current study used fMRI to measure brain responses to spoken words while performing two different listening tasks: an active listening task, in which participants listened to words while placed in the scanner, and a word repetition task, in which participants repeated words they heard aloud while in the scanner. Participants in the study were 29 young adults (aged 19-30) and 32 older adults (aged 65-81) with no self-reported hearing difficulty. For both tasks results revealed young and older adults had reliable activation in bilateral superior temporal gyrus in response to words, with an expected increase in premotor cortex and supplemental motor area in the word repetition task. Age differences were found across both tasks in auditory cortex, where older adults had significantly less activation than young adults. These age-related decrements in auditory cortex activation were selective to spoken word recognition, where no age differences were found for single-channel vocoded speech (used as a baseline auditory stimulus), and were not easily attributable to task accuracy, head movement, or audiometric hearing sensitivity (which was available on a subset of participants). This pattern of results largely suggests that older and young adults have similar patterns of brain activation in response to spoken words, albeit with a lesser degree of auditory cortex activation by older adults.

The tuning of the adult visual system to natural image statistics is evident in both behavioral and neural responses to images that do and do not violate lawful properties of natural scenes. How this sensitivity to natural image structure may develop is largely an open question however: Infants exhibit sensitivity to some properties of natural images, but adult-like sensitivity also appears to develop slowly during middle childhood. In the present study, we used event-related potentials (ERPs) to measure sensitivity to natural image statistics at components associated with early stages of visual processing (the P1/N1 complex). We hypothesized that sensitivity to violations of natural statistics (contrast negation and real vs. synthetic texture appearance) would be slow to mature during middle childhood (5-10 years of age), reflecting gradual tuning to lawful scene properties. We recruited 48 participants (5-7 year-olds, N=16; 8-10 year-olds, N=16, Adults, N=16) , who viewed natural texture images with either positive/negative contrast, and natural/synthetic texture appearance (Portilla & Simoncelli, 2000). We hypothesized that children may only acquire sensitivity to these deviations from natural texture appearance late in middle childhood (~9 years of age), consistent with previous results suggesting that texture and material processing follow a local-to-global developmental trajectory (Balas et al., 2019). We measured the P1/N1 mean amplitude and latency over midline occipital sensors in all participants. We observed significant interactions between contrast and age group for P1 latency, and between texture statistics and age group for N1 amplitude. Briefly, we found that young children exhibited sensitivity at these components to some violations of natural appearance (contrast negation, in particular) that neither older children nor adults were sensitive to. We suggest that rather than reflecting slow tuning to natural image statistics, these results may indicate that children develop broader invariant processing of natural scenes during middle childhood, effectively losing the ability to encode some distinctions that were robustly signaled earlier in development.

Early childhood is a time of rapid growth in cognitive skills, yet the accompanying changes in brain network architecture are not well understood. In this work, we examine the development of canonical functional systems in a sample of 70 children ages 4-10. We find that meso-scale measures of functional network segregation are consistently positively associated with age, including average within-network connectivity (β=0.31, p < 0.001), average between-network connectivity (β=-0.05, p < 0.001), overall system segregation (β=0.09, p = 0.026), and the average participation coefficient (β=-0.25, p = 0.008).

The relationship between age and within- and between-network connectivity varied across systems, with most significant relationships observed in connectivity between the default and attention networks. Default mode network to ventral attention network connectivity was negatively associated with age (β= -0.26, p < 0.001, pFDR = 0.015, FDR-corrected across all network comparisons), as was default mode to dorsal attention network connectivity (β= -0.21, p = 0.001, pFDR = 0.024). Visual to dorsal attention connectivity was positively associated with age (β=0.27, p = 0.007, pFDR = 0.065). We also observed positive associations between age and within-network connectivity in the visual network (β=0.26, p = 0.027), consistent with evidence that sensory networks develop earliest, as well as in the ventral attention network (β=0.27, p = 0.034), though these associations did not pass FDR correction. All analyses controlled for in-scanner motion, sex, total amount of data, and average network weight. These findings suggest increasing segregation between externally directed attention and internally oriented cognition, as well as continued development of top-down dorsal attention processes in early childhood. This work yields valuable insight into the development of functional network architecture in early childhood.

During the last few years, there has been growing interest in the effects induced by individual variability on activation patterns and brain connectivity. The practical implications of individual variability is of basic relevance for both group level and subject level studies. The Electroencephalogram (EEG), still represents one of the most used recording techniques to investigate a wide range of brain related features. In this work, we aim to estimate the effect of individual variability on a set of very simple and easily interpretable features extracted from the EEG power spectra. In particular, in an identification scenario, we investigated how the aperiodic (1/f background) component of the EEG power spectra can accurately identify subjects from a large EEG dataset. The results of this study show that the aperiodic component of the EEG signal is characterized by strong subject-specific properties, that this feature is consistent across different experimental conditions (eyes-open and eyes-closed) and outperforms the canonically-defined frequency bands. These findings suggest that the simple features (slope and offset) extracted from the aperiodic component of the EEG signal are sensitive to individual traits and may help to characterize and make inferences at single subject level.

With the aging of the population, age-related dementias are reaching epidemic levels worldwide. Alzheimer's disease (AD) is the most prevalent of the age-related dementias in adults 65 years of age and older. Traditionally AD pathology has mainly been defined by the presence of abnormal amyloid-beta and tau protein deposits. There is, however, mounting evidence that multiple factors functioning independently and/or simultaneously with amyloid-beta and tau pathologies contribute to the onset and progression of AD pathophysiological changes. Multiomics biomarkers in AD thus have the potential to reshape clinical diagnosis, and define new ‘bottom-up’ cohorts based on markers of underlying pathologies to design and evaluate drugs. We review data reduction analyses that identify complementary AD-relevant perturbations for three omics techniques: neuroimaging-based subtypes (structural and functional), metabolomics-derived metabolite panels, genomics-related polygenic risk scores. We found convergent evidence of distinct brain atrophy subtypes in AD dementia patients, including at least three data-driven atrophy subtypes. We also found that while functional subtypes of AD are emergent, the findings are in line with a recent meta-analysis (Badhwar et al, 2017). We found some evidence that AD subtypes can be characterized by metabolomics analyses. We also found that polygenic risk scores may contribute substantially to accounting for the genetic variability that distinguishes AD from mild cognitively impaired and cognitively normal groups. Following this focused review, we discuss complementarity of omics biomarkers and analytical approaches. We also present a roadmap for assembling these multiomics measurements into a diagnostic tool highly predictive of individual clinical trajectories, to further the goal of personalized medicine in AD.

Electrophysiological signals from the cerebellum have traditionally been viewed as inaccessible to magnetoencephalography (MEG) and electroencephalography (EEG). Here we challenge this position by investigating the ability of MEG and EEG to detect cerebellar activity using a model that employs a high-resolution tessellation of the cerebellar cortex. The tessellation was constructed from repetitive high-field (9.4 T) structural MRI of an ex vivo human cerebellum. A boundary-element forward model was then used to simulate the M/EEG signals resulting from neural activity in the cerebellar cortex. Despite significant signal cancellation due to the highly convoluted cerebellar cortex, we found that the cerebellar signal was on average only 30-60% weaker than the cortical signal. We also made detailed M/EEG sensitivity maps and found that MEG and EEG have highly complementary sensitivity distributions over the cerebellar cortex (Figure 1). Based on previous fMRI studies combined with our M/EEG sensitivity maps, we discuss experimental paradigms that are likely to offer high M/EEG sensitivity to cerebellar activity. Taken together, these results show that cerebellar activity should be clearly detectable by current M/EEG systems with an appropriate experimental setup.

Figure 1. Sensitivity maps of the cerebellum and the cerebral cortex for MEG magnetometers (left) and EEG (right). The color of the source corresponds to the Euclidean norm of the signal in sensor space resulting from a unit dipole of strength 100 nAm activated at that source point. The color scale was chosen to range from the 1st percentile to the 99th percentile of all cerebellar signal norms. Source points that were excluded due to the 5 mm distance limit to the inner skull boundary are black. Plane 1 is the midsagittal section of the cerebellum (vermis) and plane 2 makes a 45 degree angle with the midsagittal plane in the lateral direction.

Neuroscience research has the potential to unlock innovation in health and wellbeing, and to deepen our understanding of the human –and animal– brain and mind. Data at large scale and from multiple modalities are necessary to synthesise knowledge across multiple dimensions. Studies aggregate large and heterogeneous datasets that range from simple text files to complex hierarchical, multidimensional and multimodal data formats. It is difficult –but not impossible– to harmonize the way the data is stored and shared. Improving the interoperability of brain imaging datasets in coordination with the FAIR principles is fundamental to preventing inefficient use of human resources and delivering on the transformational and translational potential from human brain imaging.

The Brain Imaging Data Structure (BIDS) is a common standard for organizing, describing and sharing neuroimaging data. BIDS is based on simple file formats (often text-based) and folder structures that can readily expand to additional data modalities and applications. Currently it supports numerous neuroimaging modalities: magnetic resonance imaging (Gorgolewski et al., 2016), magnetoencephalography (Niso et al., 2018), electroencephalography (Pernet et al., 2019), intracranial electrophysiology (Holdgraf et al., 2019), and will soon support positron emission tomography (Knudsen et al., 2020). Multiple applications and converters have been released (Gorgolewski et al., 2017) to make it easy for researchers to incorporate BIDS into their current workflows and to maximise data sharing opportunities (for example through OpenNeuro, https://openneuro.org).

BIDS is a community-led standard. The people building these guidelines follow three foundational principles:
To minimize complexity and facilitate adoption, reuse existing methods and technologies whenever possible.
Tackle 80% of the most commonly used neuroimaging data, derivatives, and models (inspired by the pareto principle).
Adoption by the global neuroimaging community and their input during the creation of the specification is critical for the success of the project.
Following a community consultation in 2019, BIDS introduced a new leadership structure and governance system (https://bids.neuroimaging.io/governance). In this talk, chair of the BIDS Steering Group, Guiomar Niso, will share the group’s vision for the next stage of development of BIDS. The current focus includes clearer communication pathways, building resilience in the growth and maintenance of the standard, and expanding the BIDS user community. The Steering Group is committed to listening to the needs of the neuroimaging community, and all OHBMX participants will be invited to share their priorities during the talk discussion.

For a more detailed description of the BIDS specification, example datasets, resources and feedback, please visit http://bids.neuroimaging.io and follow the project on twitter at @BIDSstandard.

The OHBM Student and Postdoc SIG provides a community of shared interests for OHBM students and postdocs (trainees). Our platform provides support and network for trainees by promoting opportunities for professional, personal and career development. Additionally, we represent the needs of trainees to the OHBM Council and the wider organisation. We also promote the achievements of trainees within the organisation and to the wider neuroimaging community. The SIG organises the Monday Night Social and career development events at the annual meeting, as well as running the online, international mentoring programme.

Insight into the etiology of sex differences in the healthy brain provides a foundation with which to delineate sex-specific pathophysiology in neurological disorders displaying sex differences in clinical profiles and guide the development of sex-specific treatment. The structural connectome (SC) represents white matter tracts connecting brain regions, and the functional connectome (FC) represents similarity of electro-chemical activation over time within brain regions. Although structure and function are inexorably linked, studies thus far have analysed the networks separately. A hybrid connectome (HC) comprised of the FC’s upper triangular and SC’s lower triangular can be used to integrate the two networks and analyse them simultaneously. The marriage of the thriving field of brain connectivity network analysis and machine learning provides a promising data-driven framework with which to elucidate sex differences that exist in multimodal connectivity. Here, we use FC, SC, and HC from brain volume-matched male-female pairs (n=364) to predict sex and identify features most predictive of sex using a linear support vector machine. We find sex classification using the HC (area under the receiver operating characteristics curve, AUC=0.98), significantly outperforms (p<0.05) the independent use of FC (AUC=0.94), and SC (AUC=0.95). Analysis of the most important FC and SC features for the prediction shows a striking complementarity (r=-0.11, p<0.05). FC in frontal and occipital lobes and cerebellum are particularly important, while SC in parietal and temporal lobes and sub-cortical areas are most important. The finding that HC outperforms the independent use of FC or SC for sex prediction, and regional feature importance of FC and SC are orthogonal to one another, demonstrates that FC- and SC- based predictions are complementary. This suggests that the integration of multimodal data is crucial to understanding neurophysiological sex differences in the brain.

Brain-computer interfaces (BCIs) are devices that directly translate brain activity into commands, enabling user to interact with the world with their mind. Their main application scope is as assistive devices, to allow people with severe disabilities to communicate or operate actuators or different kinds. However, BCIs have also been used to enhance human performance in a range of cognitive tasks, including perceptual decision-making. More recently, BCIs have also been applied to groups of individuals (collaborative BCIs) to enhance critical decision-making, such as target identification from static pictures. These BCIs monitor electroencephalographic (EEG) signals of each decision-maker and use machine learning to assess how confident each person is in making his/her decision. These confidence estimates are then used to weigh individual responses and obtain a group decision.
An alternative approach to augment decision-making would be to let machines make decisions. Advances in machine learning and pattern recognition have boosted the accuracy in several domains. However, in less-controlled, realistic environments, machine vision algorithms could also fail.
This study explores the possibility of combining a residual neural network (ResNet), BCIs and human participants to improve face recognition in crowded environments. Ten human participants and a ResNet undertook the same face-recognition experiment, where a picture from a crowded scene was presented for 300 ms. Human and artificial agents were then asked to decide whether a particular face appeared on the scene. The experiment consisted of 6 blocks of 48 trials each. After data collection, BCIs were used to decode the decision confidence of humans from their EEG signals and reaction times. The ResNet also estimated its own confidence by computing the distance between the encodings of the face extracted from the image and the target face. Groups decisions were obtained weighing individual decisions by confidence estimates. Different types of groups were created, including either only humans (with or without the BCI) or groups of humans and the ResNet.
Taken individually, the accuracy of the average participant was 72.3%, while the ResNet had an average accuracy of 84.4%. The ResNet performance were characterized by high specificity (98.6%) and low sensitivity (41.7%), while the average participant had lower specificity (77.4%) and higher sensitivity (56.9%). Groups of BCI-assisted humans and ResNet were significantly more accurate (up to 35%) than the ResNet alone, the average participant, and equally-sized groups of humans not assisted by the BCI.
These results suggest that melding humans, BCI, and machine-vision technology may provide the best performance in critical decision-making in realistic scenarios.

Subject motion during fMRI can substantially affect statistical analyses and resulting interpretation. Recent years have seen increasing awareness of this fact in resting state fMRI. However, there have been fewer systematic investigations in the context of task-related fMRI. One reason for this is that, unless motion is systematically related to task timing (for example, if subjects move more during difficult conditions), it might be assumed that motion is unlikely to systematically bias results when averaged across participants. However, an increasing interest in single-subject reproducibility motivates investigating the impact of motion on reproducible results. Secondly, motion will affect the degree to which motion-free statistical models can fit the data, and potentially the accuracy of the resulting parameter estimates. We set out to systematically evaluate approaches to accounting for motion in task-based fMRI data using reproducible workflows and open data. We focused on motion “scrubbing” (that is, modeling out bad scans within the general linear model using nuisance regressors) across a range of thresholds, compared to including motion regressors or a wavelet despiking approach. Across several publicly-available datasets we found consistent numeric improvements for modest amounts of scrubbing (e.g., 1–2% data loss). However, in datasets with higher overall movement (children and older adults) we saw more marked improvements from motion scrubbing. These findings suggest motion scrubbing is frequently beneficial in task-based fMRI, and introduce a framework where metrics such as test-retest reliability might be used to assess effectiveness of this approach in new datasets.

Multivoxel pattern analysis (MVPA) allows investigators to detect information in activity collected through functional magnetic resonance imaging (fMRI) with a level of specificity that is often not possible from univariate responses. In parallel, measuring functional connectivity between regions allows us to quantify the strength of shared fluctuations in activation over time, to help us understand how regions operate in coordinated and interacting networks. Methods that combine MVPA with connectivity can bring us the best of both worlds – helping identify how brain regions interact as they process percepts and cognitive states at very fine-grained level. In the seven years since I introduced a method for combining MVPA and functional connectivity, informational connectivity (IC), a number of independent research groups have applied the technique to answer questions about how MVP information is represented in the context of other brain regions. In this presentation, I will first introduce the method, followed by separate examples of IC identified in papers and preprints that comes from different groups and areas of cognitive neuroscience. Finally, I will highlight shared broad findings and ideas to help identify common principles of neural organization that cut across subfields.

One of the challenges in Multiple Sclerosis (MS) is that the correlation between the clinical impairment and disease burden measured with structural Magnetic Resonance Imaging (MRI) is poor. Therefore, brain’s static and dynamic functional connectivity (sFC and dFC) and structural connectivity (SC) may enable a deeper understanding of the connectome-level mechanism contributing to variability in MS-related impairments. Previous studies have used various statistical methods applied to MS patient data to classify MS patients vs healthy controls (HC), classify disease phenotype, or to predict longitudinal change in impairment. However, no study to date has performed a rigorous analysis of the relative contributions of multi-modal imaging data including the sFC, dFC and SC in the task of classifying MS patients with a mild v and non-mild impairment. We aim to identify the best combination of sFC, dFC, and SC networks that can accurately classify MS patients into two groups: mild impairment (Expanded Disability Status Score (EDSS) < 2) and non-mild impairment (EDSS ≥ 2), and MS vs HC using machine learning-based models, and to investigate which particular brain connections between two regions are most associated to the motor impairment in MS. Seventy-six MS patients (age: 45.37 ± 11.44 years, 66% female, disease duration: 12.48 ± 7.22 years) were included in our study; 23 non-mild impairment at study baseline. 15 age and gender matched HC were used. Three different datasets were used in the models: regional pair-wise SC and static FC datasets, and dFC extracted from hard-clustering analysis and fuzzy meta-state analysis. Four ensemble models were built by averaging the individual probability predictions of the models that included three datasets separately. Random Forest (RF) was applied to classify MS patients as mild and non-mild impaired and MS vs HC. The performance of the models was assessed with the area under the ROC curve (AUC). The ensemble model that averaged the predictions from SC and static FC performed significantly better compared to other models in classifying MS and HC. However, the ensemble model that included SC and dynamic FC gave significantly highest AUC results. SC between left paracentral and right precuneus and FC between left frontal pole and right paracentral were the most important predictors in classifying MS patients regarding their impairment level. In conclusion, we observed that SC and static FC was the most important imaging feature to classifying MS and HC, while dynamic FC played the most important role in distinguishing MS regarding their impairment level. The connections that distinguished MS patients regarding their impairment level were between two distinct hemispheres. The ensemble model approach that was applied with RF method has the potential to help clinicians to predict the clinical impairment in MS patients, thus providing further information for personalized treatment decisions.

Encoding models based on deep neural networks (DCNN) more accurately predict BOLD responses to natural scenes in the visual system than any other currently available model. However, DCNN-based encoding models fail to predict a significant amount of variance in the activity of most voxels in all visual areas. This failure could reflect limitations in the data (e.g., a noise ceiling), or could reflect limitations of the DCNN as a model of computation in the brain. Understanding the source and structure of the unexplained variance could therefore provide helpful clues for improving models of brain computation. Here, we characterize the structure of the variance that DCNN-based encoding models cannot explain. We determined if the source of unexplained variance was shared across voxels, individual brains, retinotopic locations, and hierarchically distant visual brain areas. We answered these questions using voxel-to-voxel (vox2vox) models that predict activity in a target voxel given activity in a population of source voxels. We found that simple linear vox2vox models increased within-subject prediction accuracy over DCNN-based models for any pair of source/target visual areas, clearly demonstrating that the source of unexplained variance is widely shared within and across visual brain areas. However, vox2vox models were not more accurate than DCNN-based models when source and target voxels came from separate brains, demonstrating that the source of unexplained variance was not shared across brains. Furthermore, the weights of these vox2vox models permitted explicit readout of the receptive field location of target voxels, demonstrating that the source of unexplained variance induces correlations primarily between the activities of voxels with overlapping receptive fields. Given these results, we argue that the structured variance unexplained by DCNN-based encoding models is unlikely to be entirely caused by spatially correlated noise or eye movements; rather, our results point to a need for brain models that include endogenous dynamics and a pattern of connectivity that is not strictly feed-forward.

Pooling fMRI data across individuals requires functional correspondence across brains. However, identifying functional regions of interest (fROIs) cannot rely on “common space” anatomical alignment because, across subjects, the same voxel often belongs to functionally distinct regions. Still, in many subfields the default method remains a voxel-wise random-effects analysis (RFX). An existing alternative is a group-constrained, subject-specific method (GSS): an algorithm identifies common activation topographies across subjects, partitions this common landscape, and establishes boundaries (“masks”) within which subject-specific activations are considered functionally “the same” despite inter-individual variability in precise location. GSS has a firm mathematical foundation, but it has not been directly and comprehensively contrasted with RFX.
We compared GSS to RFX using two extensively validated tasks: a spatial working memory (sWM) task (n=291 participants), where a contrast between a harder and an easier condition robustly activates the domain-general, bilateral fronto-parietal / cingulo-opercular “executive” network; and a language task (n=352), where a contrast between passive reading of sentences vs. lists of nonwords robustly activates the domain-specific, left-lateralized fronto-temporal language network. For each task, we first partitioned a group-based probabilistic overlap map into broad masks surrounding local maxima. Within each mask, we defined a group-based RFX fROI (based on voxel-wise group statistics) or subject-specific GSS fROIs (based on statistics of individual activations). Response profiles of both fROI types were examined in held-out data using across-runs cross-validation.
Across fROIs, contrast effects obtained with RFX were at least 0.7 SDs (sWM) or 1.15 SDs (language) weaker than those obtained with GSS. Moreover, for sWM, in all but one fROI the effect of the hard condition (expected to be strong) obtained with RFX was weaker than the effect of the easy condition (expected to be weak) obtained with GSS. For language, in the anterior temporal mask, RFX did not find significant voxels despite the large sample, but GSS identified a robust contrast effect. In frontal fROIs, the RFX-based contrast effect was not reliable in ~20% of participants, who showed a numerically opposite effect in held-out data. Still, in 80% of those participants, the cross-validated GSS-based contrast effect was numerically in the right direction, thus identifying the response of interest where RFX had failed.
GSS substantially and invariably outperforms the group-based RFX in identifying fROIs. It detects truer effect sizes (increased validity); can confer internal stability via cross validation (increased reliability); and is more likely to detect effects when present (increased power). GSS is easy to implement via a downloadable toolbox, and can be applied to datasets originally designed for RFX analyses, even after data have been collected.

Background. Recent work using simultaneous EEG-fMRI recordings has shown the involvement of two major neural circuits involved in emotion-cognition interactions. These include dorsolateral prefrontal cortex (dlPFC) and ventrolateral prefrontal cortex (vlPFC). While several task-based studies have outlined the neural circuitry associated with these two regions, their intrinsic connectivity is not fully clear. Moreover, the connectivity of these networks relates to the individual differences in attention processing is unknown. In this study, we employed a simultaneous EEG-fMRI protocol to further characterize the intrinsic connectivity of dlPFC and vlPFC in healthy controls and examined how the spontaneous resting-state connectivity of the two circuits relate to individual differences in emotion and cognitive processing.
Methods. 19 healthy adults (18-31 years old, 13 females) completed the behavioural measures of Emotional Contagion Scale (ECS), and Emotional Regulation Questionnaire (ERQ), which was followed by a 5-minute resting-state scan. CONN toolbox was used to analyze the resting-state data to examine the resting-state connectivity for bilateral dlPFC and vlPFC. Finally, the ECS and ERQ scores were correlated with connectivity patterns of dlPFC and vlPFC.
Results. As expected, dlPFC was intrinsically connected to areas related to executive functioning (e.g. lateral parietal cortex), while vlPFC displayed spontaneous connections with areas associated with emotional processing (e.g. occipital-temporal cortex). Only the right dlPFC was found to be negatively correlated with the medial prefrontal cortex (part of the default mode network), indicating the functional significance of the dlPFC network for attention processing. Finally, the connectivity of the dlPFC and vlPFC related to emotional contagion and regulation behavior in an opposing fashion. vlPFC connectivity was positively related to ECS and negatively connected to ERQ, while the dlPFC displayed the opposing relationship (negatively related to ECS and positively related to ERQ).
Conclusion. This study provides emerging evidence for the presence of distinct but separate spontaneous networks for the two prefrontal cortices associated with emotion-cognition interactions. It is critical to understand the brain circuitry responsible for emotion-attention interaction to better understand how this circuitry is impaired in mood and anxiety disorders.