This axis designs and evaluates methods dedicated to exploitation of multimodal 3D-4D images for knowledge extraction in oncology applications (including lung, liver, cervix, esophagus, head and neck, rectum, and brain tumors), with a focus on the 3 following themes:

• Low data regimes (i.e., rare cancer, small cohorts specific to sub-types of cancer or patient types, as well as limited available annotations)

• Generative AI / synthesis methods

• Interpretability/explainability

Methodological developments focus on:

• Exploring alternative machine/deep learning (ML/DL) models and associated training approaches better suited small medical imaging oncological datasets, and the development of foundation models

• Incorporating a priori information (e.g. geometrical or clinical) and constraints  into ML/DL models

• Generative AI (image synthesis) for e.g., cross-modality or within-modality transforms

• Interpretability and explainability techniques dedicated to ML/DL for increased trust and reliance by clinician end-users

These focuses are then derived and implemented into the following downstream tasks:

• Image enhancement

• Tumors, organs or structures detection and segmentation

• Longitudinal or dynamic images (handling time dimension)

• Multicentric harmonization

• Radiomics and dosiomics (including combination with other types of data)

The targeted clinical applications are:

• Diagnosis/screening (e.g., differentiate between different tumor subtypes, HPV status, etc.)

• Surgery planning and post-operative evaluation

• Follow-up monitoring and prediction of outcomes

Current Projects

• 2025-2028, Thèse co-tutelle internationale UBO et Université Libanaise, Développement de modèles prédictifs explicables dans des cohortes multicentriques en oncologie
• 2024-2027, Thèse CIFRE Dosisoft, Dosiomique optimisée pour l’assurance qualité des traitements adaptatifs en radiothérapie externe
• 2023-2026, Thèse co-tutelle internationale (Ligue contre le cancer et Brest Métropole), "Novel spatial neural networks for explainable and efficient radiomics in oncology"
   

This axis develops new methodologies for tomographic image reconstruction in PET and CT. Our methodological developments include:

• Fast image reconstruction algorithms.

• Unsupervised learning for image priors.

This techniques are developed for low-dose and low-statistic imaging, with applications including:

CT-less PET reconstruction

• Joint estimation of the activity and the attenuation from emission data only
• Utilisation of diffusion models for ultra low-dose imaging (Phung et al., 2025)

Motion-compensated CT

• Joint image reconstruction and motion estimation
• Utilisation of diffusion models forlow-dose  sparse-view CT (De Paepe et al., 2025)

Spectral CT material decomposition

• One-step material decomposition from spectral data

• Utilisation of diffusion models for ultra low-dose imaging (Vazia et al., 2025)

Other research areas

• 3-gamma PET reconstruction

• Positron range correction

• Scatter correction

Current Projects

• 2024-2027, PhD Thesis, DeepMove – deep learning for motion-compensated CT reconstruction
• 2024-2027, PhD Thesis, Motion-corrected Ultra low-dose PET with Diffusion models
• 2023-2027, PhD Thesis, Mat-DPS – Material decomposition with diffusion posterior sampling
• 2020-2025, National Collaboration, MultiRecon – Machine Learning for Multimodal Medical Image Reconstruction