2.1 Context

ARAMIS is an Inria project-team within the Paris Brain Institute (ICM) at the Pitié-Salpêtrière hospital (AP-HP) in Paris. ARAMIS was created as a team of the Inria Paris Center in 2012 and became a project-team in 2014. ARAMIS has a joint affiliation to Inria, CNRS, Inserm and Sorbonne University.

The Pitié-Salpêtrière hospital is the largest adult hospital in Europe. It is a leading center for neurological diseases: in terms of size (around 20,000 neurological patients each year), level of clinical expertise and quality of the technical facilities. Created in 2010, the Paris Brain Institute (ICM) gathers all research activities in neuroscience and neurology of the Pitié-Salpêtrière hospital. The ICM is both a private foundation and a public research unit (affiliated to CNRS, Inserm and Sorbonne University). It hosts about 25 research teams as well as various high level technical facilities (neuroimaging, genotyping/sequencing, cell culture, cellular imaging, bioinformatics ...), and gathers over 700 personnel. In addition, the ICM hosts one of the six IHU (Instituts Hospitalo-Universitaires).

ARAMIS is thus located both within a leading neuroscience institute and within a large hospital. This unique position has several advantages: direct contact with neuroscientists and clinicians allows us to foresee the emergence of new problems and opportunities for new methodological developments, provides access to unique datasets, and eases the transfer of our results to clinical research and clinical practice.

2.2 General aim

The ARAMIS team is devoted to the design of computational, mathematical and statistical approaches for the analysis of multimodal patient data, with an emphasis on neuroimaging data. The core methodological domains of our team are: machine learning, statistical modeling of complex geometric data, connectivity and network analysis. These new approaches are applied to clinical research in neurological diseases in collaboration with other teams of the ICM, clinical departments of the Pitié-Salpêtrière hospital and external partners.

We develop various clinical applications of our research, in particular in neurodegenerative disorders (Alzheimer's disease and other dementias, Parkinson's disease...), multiple sclerosis, developmental disorders, stroke and to design brain-computer interfaces for rehabilitation.

3.1 Neuroimaging-based biomarkers and decision support systems

Neuroimaging provides critical information on anatomical and functional alterations as well as on specific molecular and cellular processes. Our work is focused on the development of computational approaches to extract biomarkers and build computer-aided diagnosis (CAD) systems from MRI and PET data. More specifically, we developed: i) image translation models that can generate biomarkers of specific pathological processes from unspecific routine imaging data; ii) approaches for detecting local abnormalities; iii) frameworks for reproducible and reliable evaluation of CAD systems; iv) methods for training and validating from large-scale hospital data warehouses.

3.2 Models of brain networks

Functional imaging techniques (EEG, MEG and fMRI) allow characterizing the statistical interactions between the activities of different brain areas, i.e. functional connectivity. Functional integration of spatially distributed brain regions is a well-known mechanism underlying various cognitive tasks, and is disrupted in brain disorders. Our team develops mathematical frameworks to analyze and model brain networks, or graphs, estimated from experimental data via network-science approaches. More specifically, we proposed analytical tools to infer brain networks, characterize their structure and dynamics, over multiple spatial and temporal scales.

3.3 Disease progression modeling with longitudinal data

Longitudinal data sets contain observations of multiple subjects observed at multiple time-points. They offer a unique opportunity to understand temporal processes such as ageing or disease progression. We aim here to develop a new generation of statistical methods to infer the dynamics of changes of a series of data such as biomarkers, images or clinical endpoints, together with the variability of such multivariate trajectories within a population of reference. We apply these new models across an array of neurodegenerative diseases to i) understand the heterogeneity in disease progression, in particular how genetic factors may control variations in disease progression, ii) forecast the progression of a new patient at entry of a clinical trial for stratification purposes and iii) the design of new clinical scales for use as outcomes in trials.

3.4 High-dimensional and multimodal data

We then aim to develop tools to assist clinical decisions such as diagnosis, prognosis or inclusion in therapeutic trials. To that purpose, we leverage the tools developed by the team, such as multimodal representations, network indices and spatio-temporal models which are combined with advanced classification and regression approaches. We also dedicate strong efforts to rigorous, transparent and reproducible validation of the decision support systems on large clinical datasets.

3.5 Clinical research studies

Finally, we aim to apply advanced computational and statistical tools to clinical research studies. These studies are often performed in collaboration with other researchers of the ICM, clinicians of the Pitié-Salpêtrière hospital or external partners. Our aim is to better understand brain disorders by characterizing alterations and their progression, and to validate new tools to assist clinical decisions. While a large part of these clinical studies were in the field of dementia (Alzheimer's disease, fronto-temporal dementia), we have developed successful collaborations in other fields including multiple sclerosis, Parkinson's disease and related disorders, Huntington's disease or spino-cerebellar ataxia.

4.1 Introduction

We develop different applications of our new methodologies to brain pathologies, mainly neurodegenerative diseases. These applications aim at:

• better understanding the pathophysiology of brain disorders;
• designing systems to support clinical decisions such as diagnosis, prognosis and design of clinical trials;
• developing brain computer interfaces for clinical applications.

4.2 Understanding brain disorders

Computational and statistical approaches have the potential to help understand the pathophysiology of brain disorders. We first aim to contribute to better understand the relationships between pathological processes, anatomical and functional alterations, and symptoms. Moreover, within a single disease, there is an important variability between patients. The models that we develop have the potential to identify more homogeneous disease subtypes, that would constitute more adequate targets for new treatments. Finally, we aim to establish the chronology of the different types of alterations. We focus these activities on neurodegeneratives diseases: dementia (Alzheimer's disease, fronto-temporal dementia), Parkinson's disease, multiple sclerosis.

4.3 Supporting clinical decisions

We aim to design computational tools to support clinical decisions, including diagnosis, prognosis and the design of clinical trials. We design new approaches for extracting biomarkers from different types of data. Our tools have the potential to help clinicians in their diagnosis by providing automated classification that can integrate multiple types of data (clinical/cognitive tests, imaging, biomarkers). Predicting the evolution of disease in individual patients is even more difficult. We aim to develop approaches that can predict which alterations and symptoms will occur and when. Finally, new approaches are needed to select participants in clinical trials. Indeed, it is widely recognized that, to have a chance to be successful, treatments should be administered at a very early stage.

4.4 Brain computer interfaces for clinical applications

A brain computer interface (BCI) is a device aiming to decode brain activity, thus creating an alternate communication channel between a person and the external environment. BCI systems can be categorized on the basis of the classification of an induced or evoked brain activity. The central tenet of a BCI is the capability to distinguish different patterns of brain activity, each being associated to a particular intention or mental task. Hence adaptation, as well as learning, is a key component of a BCI because users must learn to modulate their brainwaves to generate distinct brain patterns. Usually, a BCI is considered a technology for people to substitute some lost functions. However, a BCI could also help in clinical rehabilitation to recover motor functions. Indeed, in current neuroscience-based rehabilitation it is recognized that protocols based on mental rehearsal of movements (like motor imagery practicing) are a way to access the motor system because they can induce an activation of sensorimotor networks that were affected by lesions. Hence, a BCI based on movement imagery can objectively monitor patients’ progress and their compliance with the protocol, monitoring that they are actually imagining movements. It also follows that feedback from such a BCI can provide patients with an early reinforcement in the critical phase when there is not yet an overt sign of movement recovery.

5.1 Awards

6.1 New software

6.2 New platforms

Platform Brain-computer interface

Participants: Marie-Constance Corsi, Arthur Desbois, Fabrizio De Vico Fallani [Correspondant].

Our team coordinates the developments of the Brain-Computer Interface (BCI) platform at the Centre EEG/MEG of the neuroimaging core facility of the ICM. Several projects, including our NETBCI NNIH/ANR and ATTACK Big-brain theory funded projects, as well as experiments by different researchers of the Institute (ANR BETAPARK Project), and the BCINET ERC Consolidator grant are currently being run. To reinforce the impact of the platform we have extended the contract of our Inria ADT engineer, who is significantly contributing to the extension of the Inria software OpenVibe allowing for new functionalities based on our methodological development on brain connectivity and develop new sofwtare for the individual calibration (A. Desbois)

7.1 Machine learning-based prediction of impulse control disorders in Parkinson's disease from clinical and genetic data

Participants: Johann Faouzi, Jean-Christophe Corvol, Olivier Colliot [Correspondant].

Impulse control disorders (ICDs) are frequent non-motor symptoms occurring during the course of Parkinson's disease (PD). The objective of this study was to estimate the predictability of the future occurrence of these disorders using longitudinal data, the first study using cross-validation and replication in an independent cohort. We used data from two longitudinal PD cohorts (training set: PPMI, Parkinson's Progression Markers Initiative; test set: DIGPD, Drug Interaction With Genes in Parkinson's Disease). We included 380 PD subjects from PPMI and 388 PD subjects from DIGPD, with at least two visits and with clinical and genetic data available, in our analyses. We trained three logistic regressions and a recurrent neural network to predict ICDs at the next visit using clinical risk factors and genetic variants previously associated with ICDs. We quantified performance using the area under the receiver operating characteristic curve (ROC AUC) and average precision. We compared these models to a trivial model predicting ICDs at the next visit with the status at the most recent visit. The recurrent neural network (PPMI: 0.85 [0.80 – 0.90], DIGPD: 0.802 [0.78 – 0.83]) was the only model to be significantly better than the trivial model (PPMI: ROC AUC = 0.75 [0.69 – 0.81]; DIGPD: 0.78 [0.75 – 0.80]) on both cohorts. We showed that ICDs in PD can be predicted with better accuracy with a recurrent neural network model than a trivial model. The improvement in terms of ROC AUC was higher on PPMI than on DIGPD data, but not clinically relevant in both cohorts. Our results indicate that machine learning methods are potentially useful for predicting ICDs, but further works are required to reach clinical relevance.

More details in 48.

7.2 Axial multi-layer perceptron architecture for automatic segmentation of choroid plexus in multiple sclerosis

Participants: Marius Schmidt-Mengin, Bruno Stankoff, Olivier Colliot [Correspondant].

Choroid plexuses (CP) are structures of the ventricles of the brain which produce most of the cerebrospinal fluid (CSF). Several postmortem and in vivo studies have pointed towards their role in the inflammatory process in multiple sclerosis (MS). Automatic segmentation of CP from MRI thus has high value for studying their characteristics in large cohorts of patients. To the best of our knowledge, the only freely available tool for CP segmentation is FreeSurfer but its accuracy for this specific structure is poor. In this paper, we propose to automatically segment CP from non-contrast enhanced T1-weighted MRI. To that end, we introduce a new model called "Axial-MLP" based on an assembly of Axial multi-layer perceptrons (MLPs). This is inspired by recent works which showed that the self-attention layers of Transformers can be replaced with MLPs. This approach is systematically compared with a standard 3D U-Net, nnU-Net, Freesurfer and FastSurfer. For our experiments, we make use of a dataset of 141 subjects (44 controls and 97 patients with MS). We show that all the tested deep learning (DL) methods outperform FreeSurfer (Dice around 0.7 for DL vs 0.33 for FreeSurfer). Axial-MLP is competitive with U-Nets even though it is slightly less accurate. The conclusions of our paper are two-fold: 1) the studied deep learning methods could be useful tools to study CP in large cohorts of MS patients; 2) Axial-MLP is a potentially viable alternative to convolutional neural networks for such tasks, although it could benefit from further improvements.

More details in 90.

7.3 Disease progression score estimation from multimodal imaging and microRNA data using supervised variational autoencoders

Participants: Virgilio Kmetzsch, Emmanuelle Becker, Isabelle Le Ber, Olivier Colliot [Correspondant].

Frontotemporal dementia and amyotrophic lateral sclerosis are rare neurodegenerative diseases with no effective treatment. The development of biomarkers allowing an accurate assessment of disease progression is crucial for evaluating new therapies. Concretely, neuroimaging and transcriptomic (microRNA) data have been shown useful in tracking their progression. However, no single biomarker can accurately measure progression in these complex diseases. Additionally, large samples are not available for such rare disorders. It is thus essential to develop methods that can model disease progression by combining multiple biomarkers from small samples. In this paper, we propose a new framework for computing a disease progression score (DPS) from cross-sectional multimodal data. Specifically, we introduce a supervised multimodal variational autoencoder that can infer a meaningful latent space, where latent representations are placed along a disease trajectory. A score is computed by orthogonal projections onto this path. We evaluate our framework with multiple synthetic datasets and with a real dataset containing 14 patients, 40 presymptomatic genetic mutation carriers and 37 controls from the PREV-DEMALS study. There is no ground truth for the DPS in real-world scenarios, therefore we use the area under the ROC curve (AUC) as a proxy metric. Results with the synthetic datasets support this choice, since the higher the AUC, the more accurate the predicted simulated DPS. Experiments with the real dataset demonstrate better performance in comparison with stateof-the-art approaches. The proposed framework thus leverages cross-sectional multimodal datasets with small sample sizes to objectively measure disease progression, with potential application in clinical trials.

More details in 54.

7.4 MicroRNA signatures in genetic frontotemporal dementia and amyotrophic lateral sclerosis

Participants: Virgilio Kmetzsch, Isabelle Le Ber, Olivier Colliot [Correspondant], Emmanuelle Becker.

MicroRNAs are promising biomarkers of frontotemporal dementia (FTD) and amyotrophic lateral sclerosis (ALS), but discrepant results between studies have so far hampered their use in clinical trials. We aim to assess all previously identified circulating microRNA signatures as potential biomarkers of genetic FTD and/or ALS, using homogeneous, independent validation cohorts of C9orf72 and GRN mutation carriers. 104 individuals carrying a C9orf72 or a GRN mutation, along with 31 controls, were recruited through the French research network on FTD/ALS. All subjects underwent blood sampling, from which circulating microRNAs were extracted. We measured differences in the expression levels of 65 microRNAs, selected from 15 published studies about FTD or ALS, between 31 controls, 17 C9orf72 presymptomatic subjects, and 29 C9orf72 patients. We also assessed differences in the expression levels of 30 microRNAs, selected from five studies about FTD, between 31 controls, 30 GRN presymptomatic subjects, and 28 GRN patients. More than half (35/65) of the selected microRNAs were differentially expressed in the C9orf72 cohort, while only a small proportion (5/30) of micro-RNAs were differentially expressed in the GRN cohort. In multivariate analyses, only individuals in the C9orf72 cohort could be adequately classified (ROC AUC up to 0.98 for controls versus presymptomatic subjects, 0.94 for controls versus patients, and 0.77 for presymptomatic subjects versus patients) with some of the signatures. Our results suggest that previously identified microRNAs using sporadic or mixed cohorts of FTD and ALS patients could potentially serve as biomarkers of C9orf72-associated disease, but not GRN-associated disease.

More details in 55.

7.5 Automatic quality control of brain T1-weighted magnetic resonance images for a clinical data warehouse

Participants: Simona Bottani, Ninon Burgos, Aurelien Maire, Didier Dormont, Olivier Colliot [Correspondant].

We published the first paper on AI for medical imaging based on the data warehouse of the Greater Paris hospitals (AP-HP). We developed and validated an automatic quality control system for brain MRI data. This is an important step to allow fast data curation and thus perform very-large scale studies using this data warehouse which comprises several million of participants. More details about the paper below.

Many studies on machine learning (ML) for computer-aided diagnosis have so far been mostly restricted to high-quality research data. Clinical data warehouses, gathering routine examinations from hospitals, offer great promises for training and validation of ML models in a realistic setting. However, the use of such clinical data warehouses requires quality control (QC) tools. Visual QC by experts is time-consuming and does not scale to large datasets. In this paper, we propose a convolutional neural network (CNN) for the automatic QC of 3D T1-weighted brain MRI for a large heterogeneous clinical data warehouse. To that purpose, we used the data warehouse of the hospitals of the Greater Paris area (Assistance Publique-Hôpitaux de Paris [AP-HP]). Specifically, the objectives were: 1) to identify images which are not proper T1-weighted brain MRIs; 2) to identify acquisitions for which gadolinium was injected; 3) to rate the overall image quality. We used 5000 images for training and validation and a separate set of 500 images for testing. In order to train/validate the CNN, the data were annotated by two trained raters according to a visual QC protocol that we specifically designed for application in the setting of a data warehouse. For objectives 1 and 2, our approach achieved excellent accuracy (balanced accuracy and F1-score > 90%), similar to the human raters. For objective 3, the performance was good but substantially lower than that of human raters. Nevertheless, the automatic approach accurately identified (balanced accuracy and F1-score > 80%) low quality images, which would typically need to be excluded. Overall, our approach shall be useful for exploiting hospital data warehouses in medical image computing.

More details in 35.

7.6 Homogenization of brain MRI from a clinical data warehouse using contrast-enhanced to non-contrast-enhanced image translation with U-Net derived models

Participants: Simona Bottani, Elina Thibeau-Sutre, Aurelien Maire, Didier Dormont, Olivier Colliot, Ninon Burgos [Correspondant].

Clinical data warehouses provide access to massive amounts of medical images and thus offer unprecedented opportunities for research. However, they also pose important challenges, a major challenge being their heterogeneity. In particular, they contain patients with numerous different diseases. The exploration of some neurological diseases with magnetic resonance imaging (MRI) requires injecting a gadolinium-based contrast agent (for instance to detect tumors or other contrast-enhancing lesions) while other diseases do not require such injection. Image harmonization is a key factor to enable unbiased differential diagnosis in such context. Additionally, classical neuroimaging software tools that extract features used as inputs of classification algorithms are typically applied only to images without gadolinium. The objective of this work is to homogenize images from a clinical data warehouse and enable the extraction of consistent features from brain MR images, no matter the initial presence or absence of gadolinium. We propose a deep learning approach based on a 3D U-Net to translate contrast-enhanced into non-contrast-enhanced T1-weighted brain MRI. The approach was trained/validated using 230 image pairs and tested on 26 image pairs of good quality and 51 image pairs of low quality from the data warehouse of the hospitals of the Greater Paris area (Assistance Publique-Hôpitaux de Paris [AP-HP]). We tested two different 3D U-Net architectures and we chose the one reaching the best image similarity metrics for a further validation for a segmentation task. We tested two 3D U-Net architectures with the addition either of residual connections or of attention mechanisms. The U-Net with attention mechanisms reached the best image similarity metrics and was further validated on a segmentation task. We showed that features extracted from the synthetic images (gray matter, white matter and cerebrospinal fluid volumes) were closer to those obtained from the non-contrast-enhanced T1-weighted brain MRI (considered as reference) than the original, contrast-enhanced, images.

More details in 80.

7.7 Multilevel survival modeling with structured penalties for disease prediction from imaging genetics data

Participants: Pascal Lu, Olivier Colliot [Correspondant].

We introduced a framework for disease prediction from multimodal genetic and imaging data. We propose a multilevel survival model which allows predicting the time of occurrence of a future disease state in patients initially exhibiting mild symptoms. This new multilevel setting allows modeling the interactions between genetic and imaging variables. This is in contrast with classical additive models which treat all modalities in the same manner and can result in undesirable elimination of specific modalities when their contributions are unbalanced. Moreover, the use of a survival model allows overcoming the limitations of previous approaches based on classification which consider a fixed time frame. Furthermore, we introduce specific penalties taking into account the structure of the different types of data, such as a group lasso penalty over the genetic modality a a l2-penalty over the imaging modality. Finally, we propose a fast optimization algorithm, based on a proximal gradient method. The approach was applied to the prediction of Alzheimer’s disease (AD) among patients with mild cognitive impairment(MCI) based on genetic (single nucleotide polymorphisms) and imaging (anatomical MRI measures) data from the ADNI database. The experiments demonstrate the effectiveness of the method for predicting the time of conversion to AD. It revealed how genetic variants and brain imaging alterations interact in the prediction of future disease status. The approach is generic and could potentially be useful for the prediction of other diseases.

More details in 58.

7.8 Data augmentation on neuroimaging data with variational autoencoders

Participants: Clément Chadebec, Elina Thibeau-Sutre, Ninon Burgos, Stéphanie Allassonnière [Correspondant].

We proposed a new method to perform data augmentation in a reliable way in the High Dimensional Low Sample Size (HDLSS) setting using a geometry-based variational autoencoder. Our approach combines a proper latent space modeling of the VAE seen as a Riemannian manifold with a new generation scheme which produces more meaningful samples especially in the context of small data sets. The proposed method is tested through a wide experimental study where its robustness to data sets, classifiers and training samples size is stressed. It is also validated on a medical imaging classification task on the challenging ADNI database where a small number of 3D brain MRIs are considered and augmented using the proposed VAE framework. In each case, the proposed method allows for a significant and reliable gain in the classification metrics. For instance, balanced accuracy jumps from 66.3% to 74.3% for a state-of-the-art CNN classifier trained with 50 MRIs of cognitively normal (CN) and 50 Alzheimer disease (AD) patients and from 77.7% to 86.3% when trained with 243 CN and 210 AD while improving greatly sensitivity and specificity metrics.

More details in 38.

7.9 ClinicaDL: an open-source deep learning software for reproducible neuroimaging processing

Participants: Elina Thibeau-Sutre, Mauricio Diaz, Ravi Hassanaly, Alexandre Routier, Didier Dormont, Olivier Colliot, Ninon Burgos [Correspondant].

As deep learning faces a reproducibility crisis and studies on deep learning applied to neuroimaging are contaminated by methodological flaws, there is an urgent need to provide a safe environment for deep learning users to help them avoid common pitfalls that will bias and discredit their results. Several tools have been proposed to help deep learning users design their framework for neuroimaging data sets. We present ClinicaDL, one of these software tools. ClinicaDL interacts with BIDS, a standard format in the neuroimaging field, and its derivatives, so it can be used with a large variety of data sets. Moreover, it checks the absence of data leakage when inferring the results of new data with trained networks, and saves all necessary information to guarantee the reproducibility of results. The combination of ClinicaDL and its companion project Clinica allows performing an end-to-end neuroimaging analysis, from the download of raw data sets to the interpretation of trained networks, including neuroimaging preprocessing, quality check, label definition, architecture search, and network training and evaluation. We implemented ClinicaDL to bring answers to three common issues encountered by deep learning users who are not always familiar with neuroimaging data: (1) the format and preprocessing of neuroimaging data sets, (2) the contamination of the evaluation procedure by data leakage and (3) a lack of reproducibility. We hope that its use by researchers will allow producing more reliable and thus valuable scientific studies in our field.

More details in 73.

7.10 MRI field strength predicts Alzheimer's disease: a case example of bias in the ADNI data set

Participants: Elina Thibeau-Sutre, Baptiste Couvy-Duchesne, Didier Dormont, Olivier Colliot, Ninon Burgos [Correspondant].

The Alzheimer's Disease Neuroimaging Initiative (ADNI) data set has been extensively used for the prediction of the progression of prodromal patients to Alzheimer's disease dementia. However, the deep learning community is not always aware of the biases that may contaminate neuroimaging data sets, which may lead to flawed results. In this case example, we demonstrated how ignoring the magnetic resonance (MR) field strength can bias performance of deep learning prediction when using MR images as input. Finally, we discussed options to overcome this problem.

More details in 91.

7.11 Pilot study of repeated blood-brain barrier disruption in patients with mild Alzheimer's disease with an implantable ultrasound device

Participants: Stéphane Epelbaum, Ninon Burgos, Alexandre Carpentier [Correspondant].

Temporary disruption of the blood-brain barrier (BBB) using pulsed ultrasound leads to the clearance of both amyloid and tau from the brain, increased neurogenesis, and mitigation of cognitive decline in pre-clinical models of Alzheimer’s disease (AD) while also increasing BBB penetration of therapeutic antibodies. The goal of this pilot clinical trial was to investigate the safety and the efficacy of this approach in patients with mild AD using an implantable ultrasound device. An implantable, 1 MHz ultrasound device (SonoCloud-1) was extradurally implanted under local anesthesia in the skull of 10 mild AD patients to target the left supra-marginal gyrus. Over 3.5 months, seven ultrasound sessions in combination with intravenous infusion of microbubbles were performed twice per month to temporarily disrupt the BBB. 18 F-Florbetapir and 18 F-fluorodeoxyglucose positron emission tomography (PET) imaging were performed on a combined PET/MRI scanner at inclusion and at four and eight months after initiation of sonications to monitor brain metabolism and amyloid levels along with cognitive evaluations. Evolution of cognitive and neuroimaging features were compared to that of a matched sample of control participants. A total of 63 BBB opening procedures were performed in nine subjects. The procedure was well-tolerated. A non-significant decrease in amyloid accumulation at four months of -6.6% (SD=7.2%) on 18 F-Florbetapir PET imaging in the sonicated gray matter targeted by the ultrasound transducer was observed compared to baseline in six subjects that completed treatments and who had evaluable imaging scans. No differences in longitudinal change in glucose metabolism were observed compared to neighboring or contralateral regions or to the change observed in the same region in ADNI participants. No significant effect on cognition evolution was observed in comparison to the ADNI participants as expected due to the small sample size and duration of the trial. These results demonstrate the safety of ultrasound-based BBB disruption and potential of this technology to be used as a therapy for AD patients. They support further research of this technique in a larger clinical trial with a device designed to sonicate larger volumes of tissue and in combination with disease modifying drugs to further enhance the effects observed.

More details in 47.

7.12 A quantified comparison of cortical atlases on the basis of trait morphometricity

Participants: Baptiste Couvy-Duchesne [Correspondant].

Many different brain atlases exist that subdivide the human cortex into dozens or hundreds of regions-of-interest (ROIs). Inconsistency across studies using one or another cortical atlas may contribute to the replication crisis across the neurosciences. Here, we provide a quantitative comparison between seven popular cortical atlases (Yeo, Desikan-Killiany, Destrieux, Julich-Brain, Gordon, Glasser, Schaefer) and vertex-wise measures (thickness, surface area, and volume), to determine which parcellation retains the most information in the analysis of behavioural traits (incl. age, sex, body mass index, and cognitive ability) in the UK Biobank sample (N 40,000). We use linear mixed models to compare whole-brain morphometricity; the proportion of trait variance accounted for when using a given atlas. Commonly-used atlases resulted in a considerable loss of information compared to vertex-wise representations of cortical structure. Morphometricity increased linearly as a function of the log-number of ROIs included in an atlas, indicating atlas-based analyses miss many true associations and yield limited prediction accuracy. Likelihood ratio tests revealed that low-dimensional atlases accounted for unique trait variance rather than variance common between atlases, suggesting that previous studies likely returned atlas-specific findings. Finally, we found that the commonly-used atlases yielded brain-behaviour associations on par with those obtained with random parcellations, where specific region boundaries were randomly generated. Our findings motivate future structural neuroimaging studies to favour vertex-wise cortical representations over coarser atlases, or to consider repeating analyses across multiple atlases, should the use of low-dimensional atlases be necessary. The insights uncovered here imply that cortical atlas choices likely contribute to the lack of reproducibility in ROI-based studies.

More details in 49.

7.13 Parsimonious model for mass-univariate vertexwise analysis

Participants: Baptiste Couvy-Duchesne [Correspondant], Olivier Colliot.

Covariance between gray-matter measurements can reflect structural or functional brain networks though it has also been shown to be influenced by confounding factors (e.g., age, head size, and scanner), which could lead to lower mapping precision (increased size of associated clusters) and create distal false positives associations in mass-univariate vertexwise analyses. We evaluated this concern by performing state-of-the-art mass-univariate analyses (general linear model, GLM) on traits simulated from real vertex-wise gray matter data (including cortical and subcortical thickness and surface area). We contrasted the results with those from linear mixed models (LMMs), which have been shown to overcome similar issues in omics association studies. We showed that when performed on a large sample ( N=8662 , UK Biobank), GLMs yielded greatly inflated false positive rate (cluster false discovery rate >0.6 ). We showed that LMMs resulted in more parsimonious results: smaller clusters and reduced false positive rate but at a cost of increased computation. Next, we performed mass-univariate association analyses on five real UKB traits (age, sex, BMI, fluid intelligence, and smoking status) and LMM yielded fewer and more localized associations. We identified 19 significant clusters displaying small associations with age, sex, and BMI, which suggest a complex architecture of at least dozens of associated areas with those phenotypes. Conclusions: The published literature could contain a large proportion of redundant (possibly confounded) associations that are largely prevented using LMMs. The parsimony of LMMs results from controlling for the joint effect of all vertices, which prevents local and distal redundant associations from reaching significance.

More details in 43

7.14 The role of critical immune genes in brain disorders: insights from neuroimaging immunogenetics

Participants: Baptiste Couvy-Duchesne [Correspondant].

Genetic variants in the human leukocyte antigen and killer cell immunoglobulin like receptor regions have been associated with many brain-related diseases, but how they shape brain structure and function remains unclear. To identify the genetic variants in HLA and KIR genes associated with human brain phenotypes, we performed a genetic association study of 30 000 European unrelated individuals using brain MRI phenotypes generated by the UK Biobank (UKB). We identified 15 HLA alleles in HLA class I and class II genes significantly associated with at least one brain MRI-based phenotypes (P < 5 × 10-8). These associations converged on several main haplotypes within the HLA. In particular, the human leukocyte antigen alleles within an ancestral haplotype 8.1 were associated with multiple MRI measures, including grey matter volume, cortical thickness (TH) and diffusion MRI (dMRI) metrics. These alleles have been strongly associated with schizophrenia. Additionally, associations were identified between HLA-DRB1*04, DQA1*03:01, DQB1*03:02 and isotropic volume fraction of diffusion MRI in multiple white matter tracts. This haplotype has been reported to be associated with Parkinson’s disease. These findings suggest shared genetic associations between brain MRI biomarkers and brain-related diseases. Additionally, we identified 169 associations between the complement component 4 (C4) gene and imaging phenotypes. We found that C4 gene copy number was associated with cortical TH and dMRI metrics. No KIR gene copy numbers were associated with image-derived phenotypes at genome-wide threshold. To address the multiple testing burden in the phenome-wide association study, we performed a multi-trait association analysis using trait-based association test that uses extended Simes procedure and identified MRI image-specific associations. This study contributes to insight into how critical immune genes affect brain-related traits as well as the development of neurological and neuropsychiatric disorders.

More details in 34

7.15 Functional connectivity ensemble method to enhance BCI performance (FUCONE)

Participants: Marie-Constance Corsi [Correspondant], Sylvain Chevallier, Fabrizio De Vico Fallani, Florian Yger.

Relying on the idea that functional connectivity provides important insights on the underlying dynamic of neuronal interactions, we propose a novel framework that combines functional connectivity estimators and covariance-based pipelines to improve the classification of mental states, such as motor imagery. A Riemannian classifier is trained for each estimator and an ensemble classifier combines the decisions in each feature space. A thorough assessment of the functional connectivity estimators is provided and the best performing pipeline among those tested, called FUCONE, is evaluated on different conditions and datasets. Using a meta-analysis to aggregate results across datasets, FUCONE performed significantly better than all state-of-the-art methods. The performance gain is mostly imputable to the improved diversity of the feature spaces, increasing the robustness of the ensemble classifier with respect to the inter- and intra-subject variability. Our results offer new insights into the need to consider functional connectivity-based methods to improve the BCI performance.

More details in 42.

7.16 Weakly supervised framework for cancer region detection of hepatocellular carcinoma in whole-slide pathologic images based on multiscale attention convolutional neural network

Participants: Songhui Diao, Yinli Tian, Wanming Hu, Jiaxin Hou, Ricardo Lambo, Zhicheng Zhang, Yaoqin Xie, Xiu Nie, Fa Zhang, Daniel Racoceanu [Correspondant], Wenjian Qin.

Visual inspection of hepatocellular carcinoma cancer regions by experienced pathologists in whole-slide images (WSIs) is a challenging, labor-intensive, and time-consuming task because of the large scale and high resolution of WSIs. Therefore, a weakly supervised framework based on a multiscale attention convolutional neural network (MSAN-CNN) was introduced into this process. Herein, patch-based images with image-level normal/tumor annotation (rather than images with pixel-level annotation) were fed into a classification neural network. To further improve the performances of cancer region detection, multiscale attention was introduced into the classification neural network. A total of 100 cases were obtained from The Cancer Genome Atlas and divided into 70 training and 30 testing data sets that were fed into the MSAN-CNN framework. The experimental results showed that this framework significantly outperforms the single-scale detection method according to the area under the curve and accuracy, sensitivity, and specificity metrics. When compared with the diagnoses made by three pathologists, MSAN-CNN performed better than a junior- and an intermediate-level pathologist, and slightly worse than a senior pathologist. Furthermore, MSAN-CNN provided a very fast detection time compared with the pathologists. Therefore, a weakly supervised framework based on MSAN-CNN has great potential to assist pathologists in the fast and accurate detection of cancer regions of hepatocellular carcinoma on WSIs.

More details in 45.

7.17 Explainability in Artificial Intelligence; towards Responsible AI

Participants: Daniel Racoceanu [Correspondant], Mehdi Ounissi, Yannick Kergosien.

Essential for a good adoption, as well as for a wise and unbiased use, explicability is a real technology lock to the evolution of Artificial Intelligence (AI), in particular concerning Machine and Deep Learning. Without an effective explicability of the proposed algorithms, these techniques will remain a black box for health (and not only) professionals, researchers, engineers and technicians - who assume (and will continue to assume) the full responsibility of their actions. Increasingly, engineers and designers of AI tools will have to demonstrate their responsibility by providing algorithms that guarantee the explicability of the proposed models. This article presents the motivations of an explainable AI, the main characteristics of the conceptual landscape of explainability in AI, the major families of explainability methods - with a focus on some of the most common methods, to finally present some of the opportunities, challenges and perspectives of this exciting field of human-machine interaction. Indeed, only through a good understanding of the challenges associated with this technological revolution that we will be able to transform AI into assets for our companies as well as for our human actors, partners and customers.

More details in 143.

7.18 Computational Pathology for Brain Disorders

Participants: Gabriel Jimenez, Daniel Racoceanu [Correspondant].

Non-invasive brain imaging techniques allow understanding the behavior and macro changes in the brain to determine the progress of a disease. However, computational pathology provides a deeper understanding of brain disorders at cellular level, able to consolidate a diagnosis and make the bridge between the medical image and the omics analysis. In traditional histopathology, histology slides are visually inspected, under the microscope, by trained pathologists. This process is time-consuming and labor-intensive; therefore, the emergence of Computational Pathology has triggered great hope to ease this tedious task and make it more robust. This chapter focuses on understanding the state-of-the-art machine learning techniques used to analyze whole slide images within the context of brain disorders. We present a selective set of remarkable machine learning algorithms providing discriminative approaches and quality results on brain disorders. These methodologies are applied to different tasks, such as monitoring mechanisms contributing to disease progression and patient survival rates, analyzing morphological phenotypes for classification and quantitative assessment of disease, improving clinical care, diagnosing tumor specimens, and intraoperative interpretation. Thanks to the recent progress in machine learning algorithms for high-content image processing, computational pathology marks the rise of a new generation of medical discoveries and clinical protocols, including in brain disorders.

7.19 Visual Deep Learning-Based Explanation for Neuritic Plaques Segmentation in Alzheimer’s Disease Using Weakly Annotated Whole Slide Histopathological Images.

Participants: Gabriel Jimenez, Anuradha Kar, Mehdi Ounissi, Lea Ingrassia, Susana Boluda, Benoit Delatour, Lev Stimmer, Daniel Racoceanu [Correspondant].

Quantifying the distribution and morphology of tau protein structures in brain tissues is key to diagnosing Alzheimer's Disease (AD) and its subtypes. Recently, deep learning (DL) models such as UNet have been successfully used for automatic segmentation of histopathological whole slide images (WSI) of biological tissues. In this study, we propose a DL-based methodology for semantic segmentation of tau lesions (i.e., neuritic plaques) in WSI of postmortem patients with AD. The state of the art in semantic segmentation of neuritic plaques in human WSI is very limited. Our study proposes a baseline able to generate a significant advantage for morphological analysis of these tauopathies for further stratification of AD patients. Essential discussions concerning biomarkers (ALZ50 versus AT8 tau antibodies), the imaging modality (different slide scanner resolutions), and the challenge of weak annotations are addressed within this seminal study. The analysis of the impact of context in plaque segmentation is important to understand the role of the micro-environment for reliable tau protein segmentation. In addition, by integrating visual interpretability, we are able to explain how the network focuses on a region of interest (ROI), giving additional insights to pathologists. Finally, the release of a new expert-annotated database and the code will be helpful for the scientific community to accelerate the development of new pipelines for human WSI processing in AD.

More details in 83.

7.20 3D reconstruction of H&E whole slide images in melanoma.

Participants: Janan Arslan, Mehdi Ounissi, Haocheng Luo, Daniel Racoceanu [Correspondant].

Cutaneous melanoma is an invasive cancer with a worldwide annual death toll of 57,000 (Arnold et al., JAMA Dermatol 2022). In metastatic state, surgical interventions are not curative and must be coupled with chemotherapy. However, even with targeted therapies, late-stage prognosis can remain poor. The complexity of melanoma stems from its tumor microenvironment; these cancer cells continually adapt to modified metabolic changes, ensuring survival and proliferation under stressful conditions. Our primary assumption is that chemotherapeutic resistance stems from a series of non-genetic transitions and changed metabolic states. Using Whole Slide Images (WSI) of melanoma tumors from Patient Derived Xenograft (PDX) mouse models, we build 3D vascular models to predict the metabolic states within the tumor. PDX samples underwent serial sectioning over 2mm depth and were stained with Hematoxylin and Eosin (H&E). Our 3D reconstruction pipeline involves three primary steps, including 2D vessel segmentation using Deep Learning, intensity and affine-based image registration, and 3D reconstruction using interpolation and 3D rendering (allowing a better interaction with biologists, pathologists, and clinicians). The originality of our computer-assisted pipeline is its capability to (a) deal with sparse data (i.e., not all tissue sections were readily available), and (b) adapt to a multitude of WSI-related challenges (e.g., epistemic uncertainty, extended processing times due to WSI scale, etc.). We posit both our 3D reconstruction pipeline, quantitative results of the major stages of the process and a detailed illustration of the challenges faced, presenting resolutions to improve the pipeline’s efficiency.

More details in 79.

7.21 A meta-graph approach for analyzing whole slide histopathological images of human brain tissue with Alzheimer disease biomarkers

Participants: Gabriel Jimenez, Pablo Mas, Anuradha Kar, Lea Ingrassia, Susana Boluda, Benoit Delatour, Lev Stimmer, Daniel Racoceanu [Correspondant].

Recently, high performance deep learning models have allowed automatic and precise analysis of high-content medical images. In digital histopathology, a typical challenge lies in analyzing whole slide images (WSI) due to their large dimensions which most often requires splitting them into small patches for feeding deep learning models. This leads to loss in global tissue level information and is particularly limiting to classification or clustering of patients based on tissue characteristics. In this study, a meta-graph approach is developed for a semantic spatial analysis of histopathological Whole Slide Images (WSI) of human brain tissue containing tau protein aggregates, one of the hallmark lesions of Alzheimer disease (AD) in brain gray matter. We propose a pipeline that extracts morphological features of tau aggregates like neuritic plaques or neurofibrillary tangles using a pre-trained U-Net model and uses these to build a graph based on Delaunay triangulation at the WSI level, in order to extract topological features from them. This pipeline is generating morphological and topological tabular data from WSI for classification and clustering patients. Further, combining locally extracted morphological features - at the neuritic plaques or neurofibrillary tangle level - with the Delaunay graph constructed at the WSI level, allows constructing a meta-graph that can be efficiently fed to graph neural network models, instead of the voluminous WSI. This pipeline is developed and tested on a dataset of 60 WSIs from various cohorts of patients having classic and rapidly advancing AD. The purpose of this pipeline is to identify novel insights into AD evolution, as well as provide a generic framework for creating knowledge rich graphs for WSI characterization and analysis.

More details in 84.

7.22 The impact of aging on human brain network target controllability

Participants: Giulia Bassignana, Olivier Colliot, Fabrizio De Vico Fallani [Correspondant].

Understanding how few distributed areas can steer large-scale brain activity is a fundamental question that has practical implications, which range from inducing specific patterns of behavior to counteracting disease. Recent endeavors based on network controllability provided fresh insights into the potential ability of single regions to influence whole brain dynamics through the underlying structural connectome. However, controlling the entire brain activity is often unfeasible and might not always be necessary. The question whether single areas can control specific target subsystems remains crucial, albeit still poorly explored. Furthermore, the structure of the brain network exhibits progressive changes across the lifespan, but little is known about the possible consequences in the controllability properties. To address these questions, we adopted a novel target controllability approach that quantifies the centrality of brain nodes in controlling specific target anatomo-functional systems. We then studied such target control centrality in human connectomes obtained from healthy individuals aged from 5 to 85. Main results showed that the sensorimotor system has a high influencing capacity, but it is difficult for other areas to influence it. Furthermore, we reported that target control centrality varies with age and that temporal-parietal regions, whose cortical thinning is crucial in dementia-related diseases, exhibit lower values in older people. By simulating targeted attacks, such as those occurring in focal stroke, we showed that the ipsilesional hemisphere is the most affected one regardless of the damaged area. Notably, such degradation in target control centrality was more evident in younger people, thus supporting early-vulnerability hypotheses after stroke.

More details in 33.

7.23 Multiscale modeling of brain network organization

Participants: Charley Presigny, Fabrizio De Vico Fallani [Correspondant].

A complete understanding of the brain requires an integrated description of the numerous scales of neural organization. It means studying the interplay of genes, synapses, and even whole brain regions which ultimately leads to different types of behavior, from perception to action, while asleep or awake. Yet, multiscale brain modeling is challenging, in part because of the difficulty to access simultaneously to information from multiple spatiotemporal scales. While some insights have been gained on the role of specific microcircuits (e.g., thalamocortical), a comprehensive characterization of how changes occurring at one scale can have an impact on other ones, remains poorly understood. Recent efforts to address this gap include the development of new analytical tools mostly adapted from network science and dynamical systems theory. These theoretical contributions provide a powerful framework to analyze and model interconnected complex systems exhibiting interactions within and between different scales, or layers. Here, we present recent advances for the characterization of the multiscale brain organization in terms of structure-function, oscillation frequencies and temporal evolution. Efforts are reviewed on the multilayer network properties underlying higher-order organization of neuronal assemblies, as well as on the identification of multimodal network-based biomarkers of brain pathologies, such as Alzheimer's disease. We conclude this Colloquium with a perspective discussion on how recent results from multilayer network theory, involving generative modeling, controllability and machine learning, could be adopted to address new questions in modern neuroscience.

More details in 67.

7.24 Temporal exponential random graph models of longitudinal brain networks after stroke

Participants: Catalina Obando, Fabrizio De Vico Fallani [Correspondant].

Plasticity after stroke is a complex phenomenon. Functional reorganization occurs not only in the perilesional tissue but throughout the brain. However, the local connection mechanisms generating such global network changes remain largely unknown. To address this question, time must be considered as a formal variable of the problem rather than a simple repeated observation. Here, we hypothesized that the presence of temporal connection motifs, such as the formation of temporal triangles ( T ) and edges ( E ) over time, would explain large-scale brain reorganization after stroke. To test our hypothesis, we adopted a statistical framework based on temporal exponential random graph models (tERGMs), where the aforementioned temporal motifs were implemented as parameters and adapted to capture global network changes after stroke. We first validated the performance on synthetic time-varying networks as compared to standard static approaches. Then, using real functional brain networks, we showed that estimates of tERGM parameters were sufficient to reproduce brain network changes from 2 weeks to 1 year after stroke. These temporal connection signatures, reflecting within-hemisphere segregation ( T ) and between hemisphere integration ( E ), were associated with patients' future behaviour. In particular, interhemispheric temporal edges significantly correlated with the chronic language and visual outcome in subcortical and cortical stroke, respectively. Our results indicate the importance of time-varying connection properties when modelling dynamic complex systems and provide fresh insights into modelling of brain network mechanisms after stroke.

More details in 65.

7.25 Safety and efficacy of riluzole in spinocerebellar ataxia type 2 in France (ATRIL): a multicentre, randomised, double-blind, placebo-controlled trial

Participants: Giulia Coarelli, Sophie Tezenas du Montcel [Correspondant], Alexandra Durr.

Riluzole has been reported to be beneficial in patients with cerebellar ataxia; however, effectiveness in individual subtypes of disease is unclear due to heterogeneity in participants’ causes and stages of disease. Our aim was to test riluzole in a single genetic disease, spinocerebellar ataxia type 2. We did a randomised, double-blind, placebo-controlled, multicentre trial (the ATRIL study) at eight national reference centres for rare diseases in France that were part of the Neurogene National Reference Centre for Rare Diseases. Participants were patients with spinocerebellar ataxia type 2 with an age at disease onset of up to 50 years and a scale for the assessment and rating of ataxia (SARA) score of at least 5 and up to 26. Patients were randomly assigned centrally (1:1) to receive either riluzole 50 mg orally or placebo twice per day for 12 months. Two visits, at baseline and at 12 months, included clinical measures and 3T brain MRI. The primary endpoint was the proportion of patients whose SARA score improved by at least 1 point. Analyses were done in the intention-to-treat population (all participants who were randomly assigned) and were done with only the observed data (complete case analysis). This trial is registered at ClinicalTrials.gov (NCT03347344) and has been completed. Between Jan 18, 2018, and June 14, 2019, we enrolled 45 patients. 22 patients were randomly assigned to receive riluzole and 23 to receive placebo. Median age was 42 years (IQR 36–57) in the riluzole group and 49 years (40–56) in the placebo group and 23 (51%) participants were women. All participants presented with moderatestage disease, characterised by a median SARA score of 13·5 (IQR 9.5–16.5). The primary endpoint, SARA score improvement of at least 1 point after 12 months, was observed in seven patients (32%) in the treated group versus nine patients (39%) in the placebo group, with a mean difference of –10.3% (95% CI –37.4% to 19.2%; p=0.75). SARA score showed a median increase (ie, worsening) of 0.5 points (IQR –1.5 to 1.5) in the riluzole group versus 0.3 points (–1.0 to 2.5) in the placebo group (p=0.70). No serious adverse event was reported in the riluzoletreated group whereas four patients in placebo group had a serious adverse event (hepatic enzyme increase, fracture of external malleolus, rectorrhagia, and depression). The number of patients with adverse events was similar in both groups (riluzole 16 [73%] patients vs placebo 19 [83%] patients; p=0.49). We were able to recruit 45 patients moderately affected by spinocerebellar ataxia type 2 for this trial. Riluzole did not improve clinical or radiological outcomes in these patients. However, our findings provide data on progression of spinocerebellar ataxia type 2 that might prove to be valuable for the design of other clinical trials.

More details in 41.

7.26 Trajectory pattern of cognitive decline in CADASIL

Participants: Sandrine Brice, Sophie Tezenas du Montcel [Correspondant], Hugues Chabriat.

The course and pattern of cognitive decline in ischemic cerebral small vessel disease remain poorly characterized. We analyzed the trajectory pattern of cognitive decline from age 25 to 75 years in cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy (CADASIL). We applied latent process mixed models to data obtained from patients with CADASIL who were repeatedly scored during their follow-up using 16 selected clinical scales or cognitive tests. The modeled evolutions of these scores obtained from 1,243 observations in 265 patients recruited at the French National Referral Centre (50.1 years on average and 45.3% men) showed wide and heterogeneous variations in amplitude along the age-related progression of the disease. Although the Backward Digit Span remained essentially stable, a linear deterioration of scores obtained using the Symbol Digit Numbers or Number of Errors of Trail Making Test B was detected from 25 to 75 years. By contrast, the largest score changes were observed at midlife using the Digit Cancellation Task. All other tests related to executive functions, memory performances, or global cognitive efficiency showed a rate of change accelerating especially at the advanced stage of the disease. Male gender and the presence of gait disorders or of some disability at baseline were found to predict earlier or large changes of 4 scores (Index of Sensitivity to Cueing, Delayed Total Recall, Initiation/Perseveration, and Barthel Index) in a subgroup of individuals distinct from the rest of the sample. Cognitive alterations develop heterogeneously during the progression of CADASIL and vary largely according to the stage of the disease. These results suggest that not only the target population and study duration but also the stage of disease progression should be considered in preparing future clinical trials aimed at reducing cognitive decline in any such condition.

More details in 36.

7.27 Temporal dynamics of the Scale for the Assessment and Rating of Ataxia in Spinocerebellar ataxias

Participants: Paul Moulaire, Pierre-Emmanuel Poulet, Sophie Tezenas du Montcel [Correspondant].

The Scale for the Assessment and Rating of Ataxia (SARA) is the reference clinical scale to assess the severity of cerebellar ataxia. In the context of upcoming therapeutic trials, a reliable clinical outcome is needed to assess the efficiency of treatments. The aim is to precisely assess and compare temporal dynamics of SARA and a new f-SARA. We analyzed data from four cohorts (EUROSCA, RISCA, CRC-SCA, and SPATAX) comprising 1210 participants and 4092 visits. The linearity of the progression and the variability were assessed using an ordinal Bayesian mixed-effect model (Leaspy). We performed sample size calculations for therapeutic trials with different scenarios to improve the responsiveness of the scale. Seven of the eight different items had a nonlinear progression. The speed of progression was different between most of the items, with an average time for a one-point increase from 3.5 years [3.4; 3.6] (median, 95% credible interval) for the fastest item to 11.4 [10.9; 12.0] years. The total SARA score had a linear progression with an average time for a one-point increase of 0.95 [0.92; 0.98] years. After removing the four last items and rescaling all items from 0 to 4, variability increased and progression was slower and thus would require a larger sample size in a future therapeutic trial. Despite a heterogeneous temporal dynamics at the item level, the global progression of SARA was linear. Changing the initial scale deteriorates the responsiveness. This new information about the temporal dynamics of the scale should help design the outcome of future clinical trials.

More details in 62.

7.28 Forecasting individual progression trajectories in Huntington disease enables more powered clinical trials

Participants: Igor Koval, Thomas Dighiero-Brecht, Sophie Tezenas du Montcel, Alexandra Durr, Stanley Durrleman [Correspondant].

Variability in neurodegenerative disease progression poses great challenges for the evaluation of potential treatments. Identifying the persons who will experience significant progression in the short term is key for the implementation of trials with smaller sample sizes. We apply here disease course mapping to forecast biomarker progression for individual carriers of the pathological CAG repeat expansions responsible for Huntington disease. We used data from two longitudinal studies (TRACK-HD and TRACK-ON) to synchronize temporal progression of 15 clinical and imaging biomarkers from 290 participants with Huntington disease. We used then the resulting HD COURSE MAP to forecast clinical endpoints from the baseline data of 11,510 participants from ENROLL-HD, an external validation cohort. We used such forecasts to select participants at risk for progression and compute the power of trials for such an enriched population. HD COURSE MAP forecasts biomarkers 5 years after the baseline measures with a maximum mean absolute error of 10 points for the total motor score and 2.15 for the total functional capacity. This allowed reducing sample sizes in trial up to 50% including participants with a higher risk for progression ensuring a more homogeneous group of participants.

More details in 56.

7.29 Progression models for imaging data with Longitudinal Variational Auto Encoders

Participants: Benoît Sauty, Stanley Durrleman [Correspondant].

Disease progression models are crucial to understanding degenerative diseases. Mixed-effects models have been consistently used to model clinical assessments or biomarkers extracted from medical images, allowing missing data imputation and prediction at any timepoint. However, such progression models have seldom been used for entire medical images. In this work, a Variational Auto Encoder is coupled with a temporal linear mixed-effect model to learn a latent representation of the data such that individual trajectories follow straight lines over time and are characterised by a few interpretable parameters. A Monte Carlo estimator is devised to iteratively optimize the networks and the statistical model. We apply this method on a synthetic data set to illustrate the disentanglement between time dependant changes and inter-subjects variability, as well as the predictive capabilities of the method. We then apply it to 3D MRI and FDG-PET data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) to recover well documented patterns of structural and metabolic alterations of the brain.

More details in 88.

7.30 Riemannian metric learning for progression modeling of longitudinal datasets

Participants: Benoît Sauty, Stanley Durrleman [Correspondant].

Explicit descriptions of the progression of biomarkers across time usually involve priors on the shapes of the trajectories. To circumvent this limitation, we propose a geometric framework to learn a manifold representation of longitudinal data. Namely, we introduce a family of Riemannian metrics that span a set of curves defined as parallel variations around a main geodesic, and apply that framework to disease progression modeling with a mixed-effects model, where the main geodesic represents the average progression of biomarkers and parallel curves describe the individual trajectories. Learning the metric from the data allows to fit the model to longitudinal datasets and provides few interpretable parameters that characterize both the group-average trajectory and individual progression profiles. Our method outperforms the 56 methods benchmarked in the TADPOLE challenge for cognitive scores prediction.

More details in 89.

7.31 Identifying health conditions associated with Alzheimer's disease up to 15 years before diagnosis: an agnostic study of French and British health records

Participants: Thomas Nedelec, Baptiste Couvy-Duschesne, Timothy Daly, Manon Ansart, Stéphane Epelbaum, Carole Dufouil, Stanley Durrleman [Correspondant].

The identification of modifiable risk factors for Alzheimer's disease is paramount for early prevention and the targeting of new interventions. We aimed to assess the associations between health conditions diagnosed in primary care and the risk of incident Alzheimer's disease over time, up to 15 years before a first Alzheimer's disease diagnosis. In this agnostic study of French and British health records, data from 20 214 patients with Alzheimer's disease in the UK and 19 458 patients with Alzheimer's disease in France were extracted from The Health Improvement Network database. We considered data recorded from Jan 1, 1996, to March 31, 2020 in the UK and from Jan 4, 1998, to Feb 20, 2019, in France. For each Alzheimer's disease case, a control was randomly assigned after matching for sex and age at last visit. We agnostically tested the associations between 123 different diagnoses of the International Classification of Diseases, 10th revision, extracted from health records, and Alzheimer's disease, by running a conditional logistic regression to account for matching of cases and controls. We focused on three time periods before diagnosis of Alzheimer's disease, to separate risk factors from early symptoms and comorbidities. Unadjusted odds ratios (ORs) and 95% CIs for the association between Alzheimer's disease and various health conditions were estimated, and p values were corrected for multiple comparisons. Depression was the first comorbid condition associated with Alzheimer's disease, appearing at least 9 years before the first clinical diagnosis, followed by anxiety, constipation, and abnormal weight loss. Interpretation These results from two independent primary care databases provide new evidence on the temporality of risk factors and early signs of Alzheimer's disease that are observable at the general practitioner level. These results could guide the implementation of new primary and secondary prevention policies.

More details in 64.

7.32 Deterministic Approximate EM Algorithm; Application to the Riemann Approximation EM and the Tempered EM

Participants: Thomas Lartigue, Stéphanie Allassonnière [Correspondant], Stanley Durrleman.

The Expectation Maximisation (EM) algorithm is widely used to optimise non-convex likelihood functions with latent variables. Many authors modified its simple design to fit more specific situations. For instance, the Expectation (E) step has been replaced by Monte Carlo (MC), Markov Chain Monte Carlo or tempered approximations, etc. Most of the well-studied approximations belong to the stochastic class. By comparison, the literature is lacking when it comes to deterministic approximations. In this paper, we introduce a theoretical framework, with state-of-the-art convergence guarantees, for any deterministic approximation of the E step. We analyse theoretically and empirically several approximations that fit into this framework. First, for intractable E-steps, we introduce a deterministic version of MC-EM using Riemann sums. A straightforward method, not requiring any hyper-parameter fine-tuning, useful when the low dimensionality does not warrant a MC-EM. Then, we consider the tempered approximation, borrowed from the Simulated Annealing literature and used to escape local extrema. We prove that the tempered EM verifies the convergence guarantees for a wider range of temperature profiles than previously considered. We showcase empirically how new non-trivial profiles can more successfully escape adversarial initialisations. Finally, we combine the Riemann and tempered approximations into a method that accomplishes both their purposes.

More details in 57.

7.33 Asymptotic Analysis of a Matrix Latent Decomposition Model

Participants: Clément Mantoux, Stéphanie Allassonnière [Correspondant], Stanley Durrleman.

Matrix data sets arise in network analysis for medical applications, where each network belongs to a subject and represents a measurable phenotype. These large dimensional data are often modeled using lower-dimensional latent variables, which explain most of the observed variability and can be used for predictive purposes. In this paper, we provide asymptotic convergence guarantees for the estimation of a hierarchical statistical model for matrix data sets. It captures the variability of matrices by modeling a truncation of their eigendecomposition. We show that this model is identifiable, and that consistent Maximum A Posteriori (MAP) estimation can be performed to estimate the distribution of eigenvalues and eigenvectors. The MAP estimator is shown to be asymptotically normal for a restricted version of the model.

More details in 59.

8.1 Bilateral grants with industry

9.1 International research visitors

9.2 European initiatives

9.3 National initiatives

10.1 Promoting scientific activities

10.2 Teaching - Supervision - Juries

10.3 Popularization

11.1 Major publications

• 1 articleM.Manon Ansart, S.Stéphane Epelbaum, G.Giulia Bassignana, A.Alexandre Bône, S.Simona Bottani, T.Tiziana Cattai, R.Raphäel Couronné, J.Johann Faouzi, I.Igor Koval, M.Maxime Louis, E.Elina Thibeau-Sutre, J.Junhao Wen, A.Adam Wild, N.Ninon Burgos, D.Didier Dormont, O.Olivier Colliot and S.Stanley Durrleman. Predicting the Progression of Mild Cognitive Impairment Using Machine Learning: A Systematic, Quantitative and Critical Review.Medical Image Analysis67January 2021, 101848
  HALDOI
• 2 articleM.Manon Ansart, S.Stéphane Epelbaum, G.Geoffroy Gagliardi, O.Olivier Colliot, D.Didier Dormont, B.Bruno Dubois, H.Harald Hampel and S.Stanley Durrleman. Reduction of recruitment costs in preclinical AD trials. Validation of automatic pre-screening algorithm for brain amyloidosis.Statistical Methods in Medical ResearchJanuary 2019, 096228021882303
  HALDOI
• 3 inproceedingsJ.Janan Arslan, H.Haocheng Luo, M.Matthieu Lacroix, P.Pierrick Dupré, P.Pawan Kumar, A.Arran Hodgkinson, S.Sarah Dandou, R. M.Romain M Larive, C.Christine Pignodel, L.Laurent Le cam, O.Ovidiu Radulescu and D.Daniel Racoceanu. 3D Reconstruction of H&E Whole Slide Images in Melanoma.SPIE Medical ImagingSPIE Medical Imaging 2023San Diego, California, United StatesFebruary 2023
  HAL
• 4 articleF.Federico Battiston, J.Jérémy Guillon, M.Mario Chavez, V.Vito Latora and F.Fabrizio De Vico Fallani. Multiplex core--periphery organization of the human connectome.Journal of the Royal Society Interface15146September 2018
  HALDOI
• 5 articleA.Alexandre Bône, O.Olivier Colliot and S.Stanley Durrleman. Learning the spatiotemporal variability in longitudinal shape data sets.International Journal of Computer VisionJuly 2020
  HALDOI
• 6 articleS.Simona Bottani, N.Ninon Burgos, A.Aurélien Maire, A.Adam Wild, S.Sébastian Ströer, D.Didier Dormont and O.Olivier Colliot. Automatic quality control of brain T1-weighted magnetic resonance images for a clinical data warehouse.Medical Image AnalysisVolume 752021
  HALDOI
• 7 articleS.Sandrine Brice, S.Sonia Reyes, A.Aude Jabouley, C.Carla Machado, C.Christina Rogan, N.Nathalie Gastellier, N.Nassira Alili, S.Stephanie Guey, E.Eric Jouvent, D.Dominique Hervé, S.Sophie Tezenas Du Montcel and H.Hugues Chabriat. Trajectory Pattern of Cognitive Decline in Cerebral Autosomal Dominant Arteriopathy With Subcortical Infarcts and Leukoencephalopathy.Neurology9910September 2022, e1019-e1031
  HALDOI
• 8 articleN.Ninon Burgos, S.Simona Bottani, J.Johann Faouzi, E.Elina Thibeau-Sutre and O.Olivier Colliot. Deep learning for brain disorders: from data processing to disease treatment.Briefings in BioinformaticsDecember 2020
  HALDOI
• 9 articleN.Ninon Burgos, J. M.Jorge M. Cardoso, J.Jorge Samper-González, M.-O.Marie-Odile Habert, S.Stanley Durrleman, S.Sébastien Ourselin and O.Olivier Colliot. Anomaly detection for the individual analysis of brain PET images.Journal of Medical Imaging802April 2021, 024003
  HALDOI
• 10 articleN.Ninon Burgos and O.Olivier Colliot. Machine learning for classification and prediction of brain diseases: recent advances and upcoming challenges.Current Opinion in Neurology3342020, 439-450
  HALDOI
• 11 articleG.Giulia Coarelli, A.Anna Heinzmann, C.Claire Ewenczyk, C.Clara Fischer, M.Marie Chupin, M.-L.Marie-Lorraine Monin, H.Hortense Hurmic, F.Fabienne Calvas, P.Patrick Calvas, C.Cyril Goizet, S.Stéphane Thobois, M.Mathieu Anheim, K.Karine Nguyen, D.David Devos, C.Christophe Verny, V. a.Vito a G Ricigliano, J.-F.Jean-François Mangin, A.Alexis Brice, S.Sophie Tezenas Du Montcel and A.Alexandra Durr. Safety and efficacy of riluzole in spinocerebellar ataxia type 2 in France (ATRIL): a multicentre, randomised, double-blind, placebo-controlled trial.The Lancet Neurology21January 2022, 225 - 233
  HALDOI
• 12 articleM.-C.Marie-Constance Corsi, M.Mario Chavez, D.Denis Schwartz, N.Nathalie George, L.Laurent Hugueville, A. E.Ari E. Kahn, S.Sophie Dupont, D.Danielle Bassett and F.Fabrizio De Vico Fallani. Functional disconnection of associative cortical areas predicts performance during BCI training.NeuroImageJanuary 2020, 116500
  HALDOI
• 13 articleB.Baptiste Couvy-Duchesne, F.Futao Zhang, K.Kathryn Kemper, J.Julia Sidorenko, N.Naomi Wray, P.Peter Visscher, O.Olivier Colliot and J.Jian Yang. Parsimonious model for mass-univariate vertexwise analysis.Journal of Medical Imaging905September 2022
  HALDOI
• 14 articleB.Baptiste Couvy‐duchesne, L. T.Lachlan T Strike, F.Futao Zhang, Y.Yan Holtz, Z.Zhili Zheng, K. E.Kathryn E Kemper, L.Loïc Yengo, O.Olivier Colliot, M.Margaret Wright, N.Naomi Wray, J.Jian Yang and P. M.Peter M. Visscher. A unified framework for association and prediction from vertex-wise grey-matter structure.Human Brain MappingJuly 2020
  HALDOI
• 15 articleF.Fabrizio De Vico Fallani and D.Danielle Bassett. Network neuroscience for optimizing brain--computer interfaces.Physics of Life Reviews31December 2019, 304-309
  HALDOI
• 16 articleF.Fabrizio De Vico Fallani, V.Vito Latora and M.Mario Chavez. A Topological Criterion for Filtering Information in Complex Brain Networks.PLoS Computational Biology131January 2017, 1-18
  HALDOI
• 17 articleS.Songhui Diao, Y.Yinli Tian, W.Wanming Hu, J.Jiaxin Hou, R.Ricardo Lambo, Z.Zhicheng Zhang, Y.Yaoqin Xie, X.Xiu Nie, F.Fa Zhang, D.Daniel Racoceanu and W.Wenjian Qin. Weakly Supervised Framework for Cancer Region Detection of Hepatocellular Carcinoma in Whole-Slide Pathologic Images Based on Multiscale Attention Convolutional Neural Network.American Journal of Pathology1923March 2022, 553-563
  HALDOI
• 18 articleA.Alexis Guyot, A. B.Ana B Graciano Fouquier, E.Emilie Gerardin, M.Marie Chupin, J.Joan Glaunès, L.Linda Marrakchi-Kacem, J.Johanne Germain, C.Claire Boutet, C.Claire Cury, L.Lucie Hertz-Pannier, A.Alexandre Vignaud, S.Stanley Durrleman, T.Thomas Henry, P.-F.Pierre-François Van De Moortele, A.Alain Trouvé and O.Olivier Colliot. A Diffeomorphic Vector Field Approach to Analyze the Thickness of the Hippocampus from 7T MRI.IEEE Transactions on Biomedical Engineering682February 2021, 393-403
  HALDOI
• 19 inproceedingsG.Gabriel Jiménez, A.Anuradha Kar, M.Mehdi Ounissi, L.Léa Ingrassia, S.Susana Boluda, B.Benoît Delatour, L.Lev Stimmer and D.Daniel Racoceanu. Visual deep learning-based explanation for neuritic plaques segmentation in Alzheimer's Disease using weakly annotated whole slide histopathological images.Lecture Notes in Computer ScienceMICCAI 2022 - 25th International Conference on Medical Image Computing and Computer Assisted InterventionLNCS-13432Medical Image Computing and Computer Assisted InterventionPart VIIISingapore, SingaporeSpringer Nature SwitzerlandSeptember 2022, 336-344
  HALDOI
• 20 articleI.Igor Koval, A.Alexandre Bône, M.Maxime Louis, T.Thomas Lartigue, S.Simona Bottani, A.Arnaud Marcoux, J.Jorge Samper-Gonzalez, N.Ninon Burgos, B.Benjamin Charlier, A.Anne Bertrand, S.Stéphane Epelbaum, O.Olivier Colliot, S.Stéphanie Allassonnière and S.Stanley Durrleman. AD Course Map charts Alzheimer’s disease progression.Scientific Reports111April 2021
  HALDOI
• 21 articleI.Igor Koval, T.Thomas Dighiero-Brecht, A. J.Allan J Tobin, S. J.Sarah J Tabrizi, R. I.Rachael I Scahill, S.Sophie Tezenas Du Montcel, S.Stanley Durrleman and A.Alexandra Durr. Forecasting individual progression trajectories in Huntington disease enables more powered clinical trials.Scientific Reports121December 2022, 18928
  HALDOI
• 22 articleP.Paul Moulaire, P.-E.Pierre-Emmanuel Poulet, E.Emilien Petit, T.Thomas Klockgether, A.Alexandra Durr, T.Tetsuo Ashizawa and S.Sophie Tezenas Du Montcel. Temporal Dynamics of the Scale for the Assessment and Rating of Ataxia in Spinocerebellar Ataxias.Movement DisordersOctober 2022
  HALDOI
• 23 articleT.Thomas Nedelec, B.Baptiste Couvy-Duchesne, F.Fleur Monnet, T.Timothy Daly, M.Manon Ansart, L.Laurène Gantzer, B.Béranger Lekens, S.Stéphane Epelbaum, C.Carole Dufouil and S.Stanley Durrleman. Identifying health conditions associated with Alzheimer's disease up to 15 years before diagnosis: an agnostic study of French and British health records.The Lancet Digital HealthMarch 2022
  HALDOI
• 24 articleD.Daniel Racoceanu, M.Mehdi Ounissi and Y. L.Yannick L. Kergosien. Explainability in Artificial Intelligence; towards Responsible AI: Instanciation dans le domaine de la santé.Techniques de l'IngenieurDecember 2022
  HALDOI
• 25 articleA.Alexandre Routier, N.Ninon Burgos, M.Mauricio Diaz, M.Michael Bacci, S.Simona Bottani, O.Omar El-Rifai, S.Sabrina Fontanella, P.Pietro Gori, J.Jérémy Guillon, A.Alexis Guyot, R.Ravi Hassanaly, T.Thomas Jacquemont, P.Pascal Lu, A.Arnaud Marcoux, T.Tristan Moreau, J.Jorge Samper-González, M.Marc Teichmann, E.Elina Thibeau-Sutre, G.Ghislain Vaillant, J.Junhao Wen, A.Adam Wild, M.-O.Marie-Odile Habert, S.Stanley Durrleman and O.Olivier Colliot. Clinica: an open source software platform for reproducible clinical neuroscience studies.Frontiers in Neuroinformatics15August 2021, 689675
  HALDOI
• 26 articleJ.Jorge Samper-González, N.Ninon Burgos, S.Simona Bottani, S.Sabrina Fontanella, P.Pascal Lu, A.Arnaud Marcoux, A.Alexandre Routier, J.Jérémy Guillon, M.Michael Bacci, J.Junhao Wen, A.Anne Bertrand, H.Hugo Bertin, M.-O.Marie-Odile Habert, S.Stanley Durrleman, T.Theodoros Evgeniou and O.Olivier Colliot. Reproducible evaluation of classification methods in Alzheimer's disease: Framework and application to MRI and PET data.NeuroImage183December 2018, 504-521
  HALDOI
• 27 articleJ.-B.Jean-Baptiste Schiratti, S.Stéphanie Allassonniere, O.Olivier Colliot and S.Stanley Durrleman. A Bayesian mixed-effects model to learn trajectories of changes from repeated manifold-valued observations.Journal of Machine Learning Research18December 2017, 1-33
  HAL
• 28 articleE.Elina Thibeau-Sutre, M.Mauricio Diaz, R.Ravi Hassanaly, A. M.Alexandre M Routier, D.Didier Dormont, O.Olivier Colliot and N.Ninon Burgos. ClinicaDL: an open-source deep learning software for reproducible neuroimaging processing.Computer Methods and Programs in Biomedicine220June 2022, 106818
  HALDOI
• 29 articleQ.Quentin Vanderbecq, E.Eric Xu, S.Sebastian Stroër, B.Baptiste Couvy-Duchesne, M.Mauricio Diaz-Melo, D.Didier Dormont and O.Olivier Colliot. Comparison and validation of seven white matter hyperintensities segmentation software in elderly patients.Neuroimage-Clinical272020, 102357
  HALDOI
• 30 articleW.Wen Wei, E.Emilie Poirion, B.Benedetta Bodini, S.Stanley Durrleman, N.Nicholas Ayache, B.Bruno Stankoff and O.Olivier Colliot. Predicting PET-derived Demyelination from Multimodal MRI using Sketcher-Refiner Adversarial Training for Multiple Sclerosis.Medical Image Analysis58101546December 2019
  HALDOI
• 31 articleJ.Junhao Wen, E.Elina Thibeau-Sutre, M.Mauricio Diaz-Melo, J.Jorge Samper-González, A.Alexandre Routier, S.Simona Bottani, D.Didier Dormont, S.Stanley Durrleman, N.Ninon Burgos and O.Olivier Colliot. Convolutional Neural Networks for Classification of Alzheimer's Disease: Overview and Reproducible Evaluation.Medical Image Analysis63July 2020, 101694
  HALDOI

11.2 Publications of the year