Thèse Intelligence Artificielle Robuste et Multimodale pour l'Irm Rapide en Contexte d'Avc Ischémique Adaptation de Domaine Biomarqueurs du Thrombus et Médecine de Précision Neurovasculaire H/F

Doctorat.Gouv.Fr

Établissement : Université Paris-Saclay GS Informatique et sciences du numérique École doctorale : Sciences et Technologies de l'Information et de la Communication Laboratoire de recherche : IBISC - Informatique, BioInformatique, Systèmes Complexes Direction de la thèse : Vincent VIGNERON ORCID 000000159176041 Début de la thèse : 2026-10-01 Date limite de candidature : 2026-08-31T23:59:59 Cette thèse s'inscrit dans la continuité du programme ACCESS/WILLIS, centré sur l'analyse automatique de l'IRM dans l'AVC ischémique aigu. Le programme a déjà structuré plusieurs briques scientifiques : ingestion des données d'IRM, harmonisation, segmentation de la lésion, détection du thrombus, estimation du mismatch, prédiction de la recanalisation, évaluation pronostique et robustesse multi-sites. Le sujet doctoral reprend ce socle mais le fait évoluer vers deux verrous plus ambitieux : la compatibilité avec les séquences IRM accélérées et la caractérisation physico-chimique du thrombus.
Le premier axe portera sur la caractérisation et la compensation du changement de domaine induit par les IRM FAST. Il s'agira de comparer, lorsque les données le permettront, des acquisitions conventionnelles et accélérées, puis d'étudier des dégradations contrôlées reproduisant les effets d'accélération : bruit, sous-résolution, pertes de contraste, flou, artefacts de reconstruction et modifications de texture. Les modèles seront évalués non seulement par des scores globaux, mais aussi par des analyses de stabilité par centre, constructeur, séquence, protocole et sous-groupe clinique. Cette approche s'inscrit dans la continuité des travaux sur la robustesse multi-site en imagerie médicale [3], mais l'applique à un contexte neurovasculaire critique où la temporalité de l'examen est directement liée au bénéfice thérapeutique.
Le second axe portera sur la modélisation du thrombus. La littérature montre que la composition du caillot influence la fibrinolyse, la facilité de thrombectomie, la résistance mécanique et le succès de la recanalisation [2,7-11]. Pourtant, cette information n'est généralement disponible qu'après extraction. L'objectif scientifique est donc d'identifier des signatures IRM pré-thérapeutiques associées à la composition du thrombus, en combinant radiomique, représentations profondes, modèles génératifs et modèles statistiques parcimonieux. Les séquences sensibles à la susceptibilité magnétique constitueront un support privilégié, mais le projet gardera une approche multimodale afin d'intégrer les informations issues de DWI, ADC, FLAIR, TOF, SWI ou SWAN.
Le troisième axe sera celui de la généralisation clinique. Les architectures modernes de segmentation médicale, telles que U-Net, 3D U-Net et nnU-Net, ont transformé l'analyse d'images biomédicales [16-18]. Les architectures transformer et hybrides, telles que Swin UNETR, permettent d'intégrer des interactions multi-échelles et des contextes plus étendus [19]. Toutefois, leur transfert clinique reste limité si les modèles apprennent des biais liés au site ou à la reconstruction. La thèse combinera donc ces architectures avec l'apprentissage invariant, l'adaptation de domaine, la quantification de l'incertitude et l'explicabilité afin de tendre vers un pipeline utilisable en conditions réelles [20-22,26-31]. Objective 1 - Robust AI models for FAST MRI
The first objective is to develop models capable of maintaining controlled performance on FAST MRI acquisitions despite changes in noise, contrast, resolution and artifacts. This objective is fundamental because clinical generalization cannot be guaranteed by performances obtained on conventional acquisitions. The main deliverable will be a robust evaluation and learning framework including degradation measures, FAST-like scenarios, site-wise validation and regularized or domain-adapted models.
Objective 2 - Non-invasive physicochemical characterization of the thrombus
The second objective is to identify candidate MRI signatures associated with thrombus composition. This is an imaging biomarker question, not a simple anatomical detection problem. The originality lies in linking in vivo imaging with ex vivo clot analyses. The expected deliverable will be an initial mapping of plausible MRI biomarkers, accompanied by stability measurements, interpretation limits and cross-validation.
Objective 3 - An explainable and generalizable clinical pipeline
The third objective is to build a coherent pipeline from data ingestion to clinical outputs, including segmentation, thrombus, mismatch, recanalization, uncertainty and explainability. The challenge is to prepare future SaMD integration. The deliverable will be a documented, reproducible, auditable scientific prototype compatible with non-regression testing. Objective 1 - Robust AI models for FAST MRI
The first objective is to develop models capable of maintaining controlled performance on FAST MRI acquisitions despite changes in noise, contrast, resolution, and artifacts. This objective is fundamental because clinical generalization cannot be guaranteed by performances obtained on conventional acquisitions. The main deliverable will be a robust evaluation and learning framework including degradation measures, FAST-like scenarios, site-wise validation, and regularized or domain-adapted models.
Objective 2 - Non-invasive physicochemical characterization of the thrombus
The second objective is to identify candidate MRI signatures associated with thrombus composition. This is an imaging biomarker question, not a simple anatomical detection problem. The originality lies in linking in vivo imaging with ex vivo clot analyses. The expected deliverable will be an initial mapping of plausible MRI biomarkers, accompanied by stability measurements, interpretation limits and cross-validation.
Objective 3 - An explainable and generalizable clinical pipeline
The third objective is to build a coherent pipeline from data ingestion to clinical outputs, including segmentation, thrombus, mismatch, recanalization, uncertainty, and explainability. The challenge is to prepare for future SaMD integration. The deliverable will be a documented, reproducible, auditable scientific prototype compatible with non-regression testing.



DEVELOPED RESEARCH QUESTIONS
RQ1 - How do FAST MRI sequences modify the statistical distributions exploited by AI models?
This question is fundamental because it directly links acquisition physics to learned representations. Its originality lies in the fact that most FAST MRI studies evaluate reconstruction quality, whereas this thesis will assess the impact on downstream tasks: segmentation, thrombus analysis, and recanalization. The bottleneck is the decomposition of effects between noise, resolution, contrast, artifacts, and reconstruction. The deliverable will be an experimental atlas of FAST domain shift, including image, representation, and clinical metrics.
RQ2 - Which physical parameters most degrade model performance?
This question is essential for moving from empirical AI to robust diagnostic AI. The originality lies in relating model errors to acquisition parameters rather than only to global scores. The bottleneck is that several factors are correlated: vendor, sequence, resolution, TE/TR, reconstruction, and motion. The deliverable will be a sensitivity analysis and a hierarchy of degradation factors by task.
RQ3 - Can representations invariant to vendors, protocols, and reconstructions be learned?
This question lies at the heart of domain adaptation and clinical generalization. It is innovative in the FAST stroke context because domains are not abstract categories but physical and organizational combinations. The bottleneck is to preserve diagnostic information while suppressing irrelevant domain signatures. The deliverable will be a comparison of invariant, adversarial, augmentation, and harmonization strategies.
RQ4 - Can reliable diagnostic information be reconstructed from accelerated acquisitions?
The question concerns not only visual reconstruction, but the preservation of information useful for decision-making. Its originality lies in placing generative models at the service of a downstream clinical task while controlling hallucination risk. The bottleneck is to distinguish useful restoration from artificial creation of structures. The deliverable will be an evaluation of generative models in terms of their impact on segmentation, thrombus analysis, mismatch, and uncertainty.
RQ5 - Are there robust MRI biomarkers associated with thrombus physicochemical composition?
This question is fundamental for neurovascular precision medicine. It is innovative because it seeks to predict latent biological properties from pre-treatment imaging signatures. The bottlenecks are the limited size of paired cohorts, the heterogeneity of ex vivo analyses, and the reliance on MRI protocols. The deliverable will be a critical list of candidate biomarkers, with robustness measures and biological hypotheses.
RQ6 - Can recanalization be predicted from multimodal thrombus signatures?
This question links imaging, clot biology, and therapeutic outcome. Its originality is to integrate the thrombus as a decision object and not merely as an anatomical obstacle. The bottleneck is probabilistic calibration with limited, heterogeneous data. The deliverable will be an exploratory probabilistic prediction model evaluated by AUC, PR-AUC, Brier score, calibration, and clinical decision curves.
RQ7 - How can physical models and deep learning be integrated in an explainable and generalizable framework?
This question goes beyond stroke and addresses scientific AI in imaging. Its originality lies in using physics not as a theoretical background but as a design, simulation, and audit constraint. The bottleneck is the compatibility between complex models, clinical explainability, and multicenter validation. The deliverable will be a physics-informed analysis framework including simulations, constraints, uncertainty, and representation audit. The method will first rely on a reproducible pipeline for ingestion and harmonization of DWI, ADC, FLAIR, TOF, SWI, SWAN or equivalent sequences. Data will be converted, anonymized, quality-controlled, aligned and documented. Acquisition metadata will be preserved whenever available in order to link performance to physical parameters and protocols. The quality of the learning database will be considered a scientific object: rate of usable cases, heterogeneity, distributions by site, missing sequences, annotation quality and exclusion conditions.
A domain-shift analysis will be conducted between conventional and accelerated acquisitions. It will combine image metrics, latent representation metrics and clinical performance metrics. The effects of resolution, noise, contrast, partial volume and reconstruction artifacts will be analyzed through stress tests and controlled degradations. Existing models will first be evaluated without adaptation in order to establish a baseline, and will then be made more robust through FAST-like augmentation, harmonization, domain adaptation, regularization and invariant learning.
Segmentation models will include 2D/3D architectures such as U-Net, nnU-Net, Swin-UNet/Swin-UNETR and multi-scale variants. Tasks will include ischemic core segmentation, thrombus detection or segmentation, mismatch estimation and feature extraction. Generative models, including VAEs and diffusion models, will be used to explore diagnostic restoration, harmonization and latent representation. Each generative use will be accompanied by an evaluation of fidelity, hallucination risk and impact on downstream tasks.
The thrombus component will combine radiomics, deep representations, parsimonious models and robust statistical analyses. The link between in vivo imaging and ex vivo characterization will be treated as a low-sample multimodal problem. Validations will use splits by site, by vendor, by protocol and, when possible, leave-one-center-out validations. Performances will be analyzed using Dice, IoU, volumetric errors, AUC, PR-AUC, Brier score, calibration, decision curves, uncertainty and difficult-case analysis.