Search

The analysis of neuroimaging data is essential for the interpretation of the functional or structural characteristics of the human brain. New machine learning algorithms usually require a high amount of data often infeasible to acquire in clinical and practical conditions. This requirement is a consequence of significant data variability arising from numerous factors (various recording procedures, subjects and sessions, presence of high levels of noise). To address this problem, in this thesis, we have investigated and proposed convolutional machine learning models adapted to the properties and well grounded assumptions about the acquired data. Therefore, the models are endowed with valuable knowledge and consequently more efficiently learn to perform certain inferences. In particular, we have studied models for the analysis of non-invasive and in-vivo structural and functional neuroimaging data, namely diffusion Magnetic Resonance Imaging (dMRI) and magneto- and electro-encephalography (M/EEG) signals.Diffusion MRI is a nuclear imaging modality which captures micro-structural properties of the examined tissue. As q-space sampling has been the most widely used high angular resolution diffusion imaging protocol (HARDI) over the last decade, we have studied spherical rotation equivariant convolutional neural networks (CNNs) for dMRI local modeling. As a first contribution, we have proposed a spherical U-net for the estimation of fiber orientation distribution functions (fODFs) with convolutions and non-linearities realized in the spectral and signal domains, respectively. To avoid aliasing, our second contribution proposes a Fourier domain CNN for micro-structure parameter estimation, where non-linearities are defined in the spectral domain.M/EEG are functional imaging techniques which measure magnetic field strength and electric field potential caused by neural electric activities in the cerebral cortex. Measured signals can be explained by Maxwell's equations with quasi-static approximations. Consequently, we can assume that cortical brain activities spread instantaneously and linearly over the measuring sensors, thus a multivariate M/EEG signal can be represented as a sum of rank-1 multivariate signals corresponding to individual sources in the cortex and noise. Considering this assumption, the second part of the thesis firstly investigates an M/EEG spatial and temporal dictionary learning approach with an $L_0$ constraint. A second contribution is a CNN classifier with rank-1 spatio-temporal kernels regularized in the spectral domain, where the spatial components of the kernels are represented in terms of spherical harmonics basis, while the temporal components are represented in terms of discrete cosine basis., L'analyse des données de neuroimagerie est essentielle pour l'interprétation des caractéristiques fonctionnelles ou structurelles du cerveau humain. Les algorithmes d’apprentissage automatique récents requièrent généralement une grande quantité de données souvent impossibles à acquérir dans des conditions cliniques et pratiques. Une telle exigence est une conséquence de la variabilité importante des données résultant de nombreux facteurs (différentes procédures d'enregistrement, sujets et sessions, présence de niveaux élevés de bruit). Pour résoudre ce problème, dans cette thèse, nous avons étudié et proposé des modèles convolutifs d'apprentissage automatique adaptés aux propriétés et aux hypothèses bien fondées sur les données acquises. Par conséquent, les modèles sont dotés de connaissances précieuses et apprennent plus efficacement à effectuer certaines inférences. En particulier, nous avons étudié des modèles d'analyse des données de neuroimagerie structurelle et fonctionnelle non-invasives et in-vivo pour de l'imagerie par résonance magnétique de diffusion (IRMd) et des signaux de magnéto et d'électro-encéphalographie (M/EEG).L'IRM de diffusion est une modalité d'imagerie nucléaire qui capture les propriétés microstructurales des tissus examinés. Comme l'échantillonnage de q-space est le protocole d'imagerie de diffusion à haute résolution angulaire (HARDI) le plus largement utilisé au cours de la dernière décennie, nous avons étudié les réseaux de neurones convolutionnels (CNN) sphériques équivariants par rotation pour la modélisation locale de l'IRMd. Comme première contribution, nous avons proposé un U-net sphérique pour l'estimation des fonctions de distribution d'orientation des fibres (fODF) avec des convolutions et des non-linéarités réalisées respectivement dans les domaines spectral et signal. Pour éviter l'aliasing, la deuxième contribution propose un CNN travaillant entièrement dans le domain spectral -- y compris pour les non-linéarités -- pour l'estimation des paramètres de microstructure.La M/EEG est une technique d'imagerie fonctionnelle qui mesure l'intensité du champ magnétique et le potentiel du champ électrique provoqués par les activités électriques neurales dans le cortex cérébral. Les signaux mesurés peuvent être expliqués par les équations de Maxwell avec des approximations quasi-statiques. Par conséquent, nous pouvons supposer que les activités cérébrales corticales se propagent instantanément et linéairement sur les capteurs de mesure, ainsi un signal M/EEG multivarié peut être représenté comme une somme de signaux multivariés de rang 1 correspondant à des sources individuelles dans le cortex et le bruit. Partant de cette hypothèse, la deuxième partie de la thèse étudie une approche d'apprentissage de dictionnaire spatio-temporel M/EEG sous contrainte $L_0$. Une deuxième contribution dans cette partie est un classificateur CNN à noyaux spatio-temporels de rang 1 régularisés dans le domaine spectral, où les composantes spatiales et temporelles des noyaux sont représenteés respectivement en termes d'éléments de base d'harmoniques sphériques et de base de cosinus discrets.

Estimating structural connectivity from diffusion-weighted magnetic resonance imaging is a challenging task, partly due to the presence of false-positive connections and the misestimation of connection weights. Building on previous efforts, the MICCAI-CDMRI Diffusion-Simulated Connectivity (DiSCo) challenge was carried out to evaluate state-of-the-art connectivity methods using novel large-scale numerical phantoms. The diffusion signal for the phantoms was obtained from Monte Carlo simulations. The results of the challenge suggest that methods selected by the 14 teams participating in the challenge can provide high correlations between estimated and ground-truth connectivity weights, in complex numerical environments. Additionally, the methods used by the participating teams were able to accurately identify the binary connectivity of the numerical dataset. However, specific false positive and false negative connections were consistently estimated across all methods. Although the challenge dataset doesn't capture the complexity of a real brain, it provided unique data with known macrostructure and microstructure ground-truth properties to facilitate the development of connectivity estimation methods.

International audience

International audience; Brain-computer interfaces (BCIs) can be separated into two main types: active and passive BCIs (Clerc et al. 2016). A BCI can be qualified of passive when the system uses signals involuntarily generated by the user. More specifically, this type of BCI is often used with the aim to assess the mental workload of users performing various task with different levels of mental demand, especially with electroencephalography (EEG) (Wang et al. 2015, Appriou et al. 2018, Shalchy et al. 2020). In most cases those systems are built with a classifier that classifies brain signals into different categories. This relies on having collected labelled data beforehand. However, those systems are often developed in laboratory settings, where both the train and test set have known labels. The "Grand Challenge: Passive BCI Hackathon" organised for the Neuroergonomics 2021 conference enables to challenge researchers with a real-life scenario of a passive BCI: classifying data from unseen sessions, with labels concealed for them, preventing any kind of fine tuning on the test set. The dataset provided for this challenge (Hinss et al. 2021) was composed of EEG recordings of 15 participants performing in 3 distinct sessions the Multi-Attribute Task Battery-II (MATB-II) developed by the NASA. Each session is decomposed in blocks of different difficulties: easy, medium and difficult. The data provided consists in epochs of 2 seconds (with a sampling frequency of 250 Hz) from those blocks for a total 447 epochs for each session and each participant. Difficulty labels were provided only for the 2 first sessions.

International audience

International audience; Diffusion Magnetic Resonance Imaging (dMRI) is an imaging technique which enables analysis of the brain tissue at a microscopic scale, particularly the analysis of white matter. Given a high enough angular resolution, a common way to explain the measured signal is via fiber orientation distribution function (fODF). This function describes the orientation and volume fraction of axon bundles within each voxel and is an essential ingredient of tractography. In this work, we have investigated a deep learning approach for the fODF estimation. U-nets enable fast and high resolution inference by combining multi-scale features from contracting and expanding parts of the network. As dMRI signals are most commonly acquired on spheres, we propose a spherical U-net which is adjusted to the properties of the dMRI data, namely its real nature, antipodal symmetry, uniform sampling and axial symmetry of the signals corresponding to individual fibers. We compared our model with another deep learning approach based on a 3D convolutional neural network and a state-of-the-art approach-multi-shell multi-tissue constrained spherical deconvolution, on real data from Human Connectome Project and synthetic data generated using ball and stick model. The methods are compared in terms of mean square error and mean angular error for dMRI signals of different angular resolutions. Provided quantitative analyses show improved performance with our approach even with significantly reduced number of parameters and results obtained on synthetic data indicate its robustness with respect to noise. Qualitative results illustrating the performance of the methods are also presented.