Thèse Pré-Entraînement Guidé par l'Anatomie et Ancré Cliniquement pour des Modèles de Fondation en Neuroimagerie H/F

Doctorat.Gouv.Fr

Établissement : Université Paris-Saclay GS Sciences de l'ingénierie et des systèmes École doctorale : Electrical, Optical, Bio-physics and Engineering Laboratoire de recherche : Construction de grands instruments pour la neuroimagerie : de l'imagerie en population aux champs magnétiques ultra-hauts Direction de la thèse : Benoit DUFUMIER ORCID 0000000282532363 Début de la thèse : 2026-09-07 Date limite de candidature : 2026-08-30T23:59:59 Les troubles psychiatriques et neurodégénératifs restent difficiles à diagnostiquer, à suivre et à traiter, car leurs effets sur l'anatomie cérébrale sont hétérogènes, progressifs et encore imparfaitement décrits par les biomarqueurs actuels. Dans des pathologies telles que les troubles du spectre de l'autisme, le trouble bipolaire, la schizophrénie, la maladie d'Alzheimer ou la maladie de Parkinson, la décision clinique repose encore largement sur des évaluations qualitatives ou sur des biomarqueurs dont l'évolution longitudinale reste incomplètement caractérisée. Ce projet vise à développer des modèles de fondation pour la neuroimagerie anatomique, capables d'apprendre des représentations robustes de la structure cérébrale et de caractériser les déviations individuelles par rapport à l'organisation cérébrale typique au cours du développement, du vieillissement et de la maladie.
Le projet s'articule autour de trois axes complémentaires. Le premier consiste à concevoir des générateurs d'IRM synthétiques guidés par l'anatomie, capables de produire des images cérébrales biologiquement plausibles et conditionnées par des variables telles que l'âge, à partir de cartes de segmentation. Ces données synthétiques serviront à enrichir les bases d'apprentissage et à améliorer le pré-entraînement auto-supervisé. Le deuxième axe étudiera les lois d'échelle et de mélange des modèles auto-supervisés, afin de quantifier l'effet de la taille des modèles, du volume de données, du budget de calcul et de la proportion de données réelles et synthétiques sur les performances. L'évaluation portera notamment sur des tâches cliniques longitudinales liées à la progression des maladies d'Alzheimer et de Parkinson. Le troisième axe développera des modèles vision-langage associant IRM cérébrales et comptes rendus cliniques, afin d'exploiter l'information contextuelle contenue dans les rapports neurologiques et psychiatriques tout en l'ancrant dans des mesures quantitatives d'imagerie.
À terme, ce projet ambitionne de produire des représentations transférables et cliniquement pertinentes de l'IRM anatomique, adaptables à des jeux de données cliniques plus restreints. Il pourrait ainsi contribuer à améliorer le diagnostic précoce, la stratification des patients, le suivi de la progression des maladies et l'interprétation biologique des troubles psychiatriques et neurodégénératifs. The biological effects of psychiatric disorders such as Autism Spectrum Disorder (ASD) remain only partially understood. As a result, their classification and diagnosis still rely largely on behavioural and qualitative criteria, with limited grounding in objective biological markers [1]. A similar challenge exists for the prodromal or preclinical stages of neurodegenerative diseases such as Alzheimer's disease (AD) and Parkinson's disease (PD), where pathological changes may begin years before clinical symptoms appear. Although several biomarkers have been identified, their longitudinal evolution and relationship to individual disease trajectories remain incompletely characterised [2, 3].
This limited understanding of disease mechanisms constrains early diagnosis, personalised treatment, patient stratification, and drug discovery, thereby contributing to the substantial human and economic burden associated with psychiatric and neurodegenerative conditions [4].
The main objective of this thesis is to characterize the clinical evolution of patients with neurodegenerative and psychiatric disorders from the prodromal phase to the first onset of the symptoms using brain imaging. To do so, we aim at building a new representation of the anatomical brain that is best predictive of this clinical evolution using Deep Learning. The idea is to (pre)train a model on an aggregation of large-scale population imaging datasets in order to capture the anatomical variability in the general population [5, 6]. The hypothesis is that there exists a low-dimensional space (or manifold) on which the brain MRI of all healthy subjects lie. Brain disorders can be represented as a deviation to this manifold [7, 8, 9]. We also hypothesize that longitudinal trajectories can be more easily captured in this low-dimensional space rather than in the original high-dimensional imaging space.


We aim at building such low-dimensional space with Self-Supervised Learning (SSL) and Vision-Language Modeling. The former does not need labels or annotations for training, but rather it relies on statistical dependencies in the brain scan itself [5]. The second relies on pairs MRI scans-clinical reports to take advantage of the rich information on patients' state contained in the neurological reports [10].
Aim 1 - Designing synthetic data generator to improve SSL pretraining on mixed synthetic and real data

Self-Supervised Learning (SSL) models (Masked Auto-Encoder [15], Contrastive Learning [16], Joint-Embedding Predictive Architecture [14]) require a tremendous amount of data to build a representation transferable to clinical downstream tasks [11, 12]. Even if data sharing initiatives such as UKB or ABCD have emerged, they are still limited to a few hundred thousands scans. In parallel, synthetic generative approaches like SynthSeg [13] offer a viable alternative to automatically augment the size and richness of the database. If successful, this would unlock the capacity of self-supervised models that shine in other domains such as natural images [17, 18]. However, SynthSeg has been originally designed for segmentation purposes and its generation module is purely handcrafted and it does not necessarily produce anatomically plausible MRI.

The first aim of this PhD is to revisit the original approach of SynthSeg to generate plausible MRI scans conditionally to biological variables (e.g., the age). The general approach relies on two key ingredients. First, we decouple the generation of the deformation field applied to the original segmentation map of an MRI with the appearance module. Second, we modify the deformation module to be conditional to age so that it becomes anatomically plausible to the eye of a brain age predictor. The generation then becomes both anatomically plausible and diverse, two key properties for downstream SSL pretraining. This is also fundamentally different from previous works [19] that rely on a purely data-driven approach to learn the generative model. The hypothesis is that purely data-driven models need more high-quality data for accurate generation than physics-guided models in which additional prior information is injected at training time.

Aim 2 - Scaling laws and mixture laws of SSL models on longitudinal trajectory benchmarks

Predicting how SSL models will improve with 1) model size 2) data size, 3) training/computation time is crucial to understand what is needed in the field. Do we need more data ? More computation time ? Which self-supervised objective (MAE, contrastive, JEPA) has the best scaling law for our clinical tasks ?
To answer these important questions, the second aim of this PhD is to derive power-law scaling relationships [20] that predict model performance as a function of model size, dataset size, and computational budget. This PhD will specifically investigate how the proportion of real and synthetic data affects the resulting scaling laws, since different ratios may lead to distinct scaling behaviors, particularly when synthetic data are of lower quality than real data. The model performance will be assessed on clinical benchmarks relevant for neurodegenerative and psychiatric disorders. In particular, the benchmarks will be focused on quantifying the progression of Alzheimer's and Parkinson's disease in a large cohort of patients (ADNI and PPMI) based on their brain MRI scans.

Aim 3 - Vision-Language Modeling with MRI and clinical text reports

Clinical reports in neurology and psychiatry provide very rich and contextual information about patients [21]. They describe the clinical symptoms, patients' history, comorbidity and medications that all inform the clinical decision. This information is complementary to brain imaging, which offers quantifiable biological information about the patients.
Linking brain imaging to clinical reports is thus crucial for better supporting the clinical decision and orienting to the best treatment. Vision-Language Models (VLM) offer a promising tool to link these modalities.
However, existing works such as ConVIRT, GLoRIA, and Clip [10, 22, 23] are mainly crafted for the radiological domain for which the reports contain a crude description of the observations in the scan. This is very different from neurology and psychiatry, for which the clinical findings are mainly based on behavior but not brain imaging alone. As such, the two modalities are misaligned [24], they contain information unique, redundant and synergistic to orient the clinical decision. Modeling all three interactions is crucial to learn a multimodal data representation predictive of the clinical outcome.

CoMM [25] is a self-supervised approach introduced to learn these interactions based on Partial Information Decomposition theory. It has been successfully applied to several multimodal tasks but its applicability to neuroimaging data remains open.
The third aim of this PhD project is to extend CoMM to neuroimaging datasets to learn a common latent space shared between brain scans and clinical reports, where additional information coming from the latter provides a more detailed representation of the patient's state. Each modality will provide a context for the interpretation of the other, allowing for a cross-validation mechanism, e.g. in the form of biological grounding offered by brain scans to qualitative criteria currently used by psychiatrists to diagnose behavioural disorders. Datasets used for pretraining SSL/VLM models

- Adolescent Brain Cognitive Development (ABCD) study (N=10k subjects) is a longitudinal (age 10 to 20) study of brain development providing multimodal neuroimaging, OMICs, and clinical data.
- Reproducible Brain Charts including (RBC) (N=6k subjects) is an aggregation of five major brain development initiatives including neuroimaging and thorough behavioral assessments.
- UK-Biobank (UKB) (N=500k subjects) is the largest biomedical database worldwide, including subjects with imaging assessed through lifestyle, environmental, genetic, and health records data.
- OpenBHB (N=5k subjects) is a large multi-site benchmark of T1-weighted brain MRI from healthy controls only, aggregating >5,000 scans from >60 sites across 10 public datasets .It spans a broad lifespan (5-88 years) and underpins challenges on brain-age prediction.
- BIND (N=39k subjects) is a multimodal brain database comprising ~1.8 million clinical MRI scans from ~39,000 subjects, linked to EEG/polysomnography and clinical metadata. It is one of the largest open-access databases to date.


Datasets used for downstream tasks

Medium size case-control (cross-sectional) datasets covering the main psychiatric disorders: Autism Spectrum Disorder (ASD), Bipolar Disorder (BD), Schizophrenia (SCZ):
- ABIDE (N~1k) of half control and patients with ASD.
- SchizConnect (N~1k) of half control and patients with SCZ.
- BSNIP (N~1k) of controls and patients with SCZ and BD.
- BIOBD (N~500) of controls and patients with BD.

Medium size longitudinal datasets covering Alzheimer's disease/Parkinson's Disease:
- ADNI (N~2k) of controls and patients with MCI and AD
- OASIS (N~1k) of controls and patients at various stage of cognitive decline
- PPMI (N~1.5k) is a large-scale, longitudinal Parkinson's disease cohort collecting clinical, imaging, genetic, omics, sensor, biomarker, and biospecimen data.

Original datasets (longitudinal) with follow-up of clinical outcome. Over the past decade, the signature team of GAIA laboratory became a key actor leading data management and analysis of large-scale psychiatric studies involving neuroimaging. This strong collaboration provides unique access to original longitudinal datasets, in particular:
- R-LiNK (N=165) focuses on identifying biomarkers for lithium response in bipolar disorder using multimodal neuroimaging, OMICs, and clinical data. The primary outcome is predicting lithium response after two years.
- PsyCARE (N=350) tracks early psychosis progression with structural and functional MRI, assessing functional outcomes (PSP scale) at 3 months and transition to psychosis at 2 years.