Informations générales
Intitulé de l'offre : Controlled Generative Models for Biomedical Image Synthesis (M/F) (H/F)
Référence : UMR7020-MARBEL-001
Nombre de Postes : 1
Lieu de travail : MARSEILLE 13
Date de publication : mardi 1 juillet 2025
Type de contrat : CDD Doctorant
Durée du contrat : 36 mois
Date de début de la thèse : 1 octobre 2025
Quotité de travail : Complet
Rémunération : 2200 gross monthly
Section(s) CN : 07 - Sciences de l'information : traitements, systèmes intégrés matériel-logiciel, robots, commandes, images, contenus, interactions, signaux et langues
Description du sujet de thèse
Scientific Context and Objectives
Recent advances in artificial intelligence, particularly in deep learning, have significantly impacted the field of image analysis and synthesis. These methods now clearly outperform traditional approaches due to their ability to learn directly from data. The biomedical domain, characterized by the availability of specialized data and expert knowledge, offers a particularly promising field of application. However, the generation of expert-annotated datasets remains limited, owing to the complexity and cost of annotation, as well as legal and ethical constraints related to data privacy and the rarity of certain pathologies.
In this context, generative models offer a promising solution for synthesizing realistic medical data, particularly in scenarios involving class imbalance or limited data availability. The ConText-GAN model, developed at the LIS laboratory [1], has already shown its potential for generating relevant synthetic biomedical images, especially for mitigating class imbalance. This model is conditioned on semantic label maps and associated texture descriptors (feature maps), allowing fine control over the geometry and visual realism of the generated content. The resulting “label-to-image” strategy enables controlled generation of high-quality and diverse synthetic images.
Research Objectives
The primary goal of this research is to explore and develop generative methods for medical image synthesis, with a strong emphasis on precise control over generated image properties — a critical requirement in biomedical applications. The methodological framework will be structured around the following axes:
1. Optimization of the ConText-GAN model: This includes improving the robustness, visual fidelity, and adaptability of the current architecture to different imaging modalities (e.g., MRI, PET, etc.).
2. Simplification of the feature map creation process: This currently time-consuming task will be addressed through automation. We will explore approaches based on prompts (textual or semantic), and automatic label generation to reduce the manual effort required.
3. Extension to diffusion-based models: To accommodate the diversity of target modalities and applications, controlled image generation with diffusion models will be investigated. These approaches aim to further enhance image realism while preserving semantic control.
4. Generative reconstruction of anatomical surfaces: A second area of investigation will focus on surface reconstruction from sets of input images using generative methods. The objective is to develop a robust surface extrapolation approach capable of inferring coherent 3D structures from 2D imaging data.
This project lies at the intersection of computer vision, statistical modeling, and biomedical image analysis. It aims to address concrete medical needs in simulation, augmented analysis, and reduction of dependency on expert-annotated datasets.
References :
[1] M.-A. Hostin, S. Attarian, D. Bendahan, M.-E. Bellemare, “ConText-GAN: using contextual texture information for realistic and controllable medical image synthesis”, 2023 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI) (IEEE BHI 2023), Oct. Pittsburgh, USA.
[2] Fortanier E, Michel CP, Hostin MA, Delmont E, Verschueren A, Guye M, Bellemare M-E, Bendahan D, Attarian S. “Quantitative muscle MRI combined with AI-based segmentation as a follow-up biomarker for ATTRv patients: A longitudinal pilot study”. Eur J Neurol. 2025; 32:e16574.
[3] E. Fortanier, M.-A. Hostin, C. Michel, E. Delmont, M. Guye, M.-E. Bellemare, S. Attarian, D. Bendahan, “Comparison of Manual vs Artificial Intelligence–Based Muscle MRI Segmentation for Evaluating Disease Progression in Patients With CMT1A”. Neurology, 2024, 103 (10), ⟨10.1212/WNL.0000000000210013⟩
Contexte de travail
The PhD will take place in Marseille, at the Laboratoire d'Informatique et des Systèmes (LIS), within the Images & Models (I&M) research team, located on the St Jérôme campus.
The I&M team specializes in image analysis, with a focus on extracting knowledge and supporting decision-making from image data. Its research activities are oriented towards several key applications: diagnosis support, preoperative planning, morphometry, behavioral analysis, and information systems.
The PhD work will be conducted in a multidisciplinary environment, involving medical applications that provide both data and a diverse framework for study. Several ongoing collaborative projects with clinicians from AP-HM (Assistance Publique - Hôpitaux de Marseille) will offer concrete application contexts — including MRI and microscopic imaging — to support and guide the methodological developments carried out during the thesis.