Candidatez avant le 19/05/2025 en cliquant sur le lien externe ci-dessus.
Abstract
[Résumé et description en français en cliquant sur le lien externe ci-dessus]
The research project for this PhD focuses on modeling information in digital images, with an emphasis on complex inverse problems such as image restoration and reconstruction, particularly in medical imaging. It combines robust mathematical approaches, such as variational calculus, with advances in deep learning—especially generative neural networks—to improve the reconstruction of medical images from spectral CT scans.
The research team, which includes experts from the Laboratoire de Mathématiques de Bretagne Atlantique (LMBA UMR CNRS 6205 – UBO, Brest, France and UBS, Brest, France) and LaTIM (UMR Inserm,
Brest), has previously supervised two PhDs: one on image reconstruction using spectral CT, and another on single-step material decomposition. These projects established a new state of the art by reducing the radiation dose patients are exposed to while producing high-quality images from limited data.
The current project focuses on generative models based on diffusion processes, which outperform GANs (Generative Adversarial Networks) in terms of stability and ease of training through the use of stochastic differential equations. These models allow the probability distribution to be conditioned on observed data, offering great flexibility for solving complex inverse problems. However, a critical challenge is the phenomenon of hallucination, where generative models produce unrealistic but statistically plausible data. In medical imaging, this can lead to misleading information, potentially compromising radiological diagnoses. This risk is heightened when data is scarce, which is often the case in low-dose radiation scenarios.
The goal of the project is to analyze and control these hallucination phenomena to ensure that combined solutions remain reliable for reconstructing medical images. The ultimate objective is to balance image quality and clinical safety for practical applications in spectral CT and other imaging domains.
This PhD position will take place within the Laboratory of Mathematics of Brittany Atlantic (LMBA UMR CNRS 6205) at the Université Bretagne Occidentale (UBO) in Brest, France. It will be co-supervised by the Université Bretagne Sud (UBS) in Vannes, France and in collaboration with the Laboratory of Medical Information Processing (LaTIM UMR INSERM 1101, Brest).
To get more details and to apply for the position, click on the external link above.
Application deadline: 05/19/2025.
