Offre en lien avec l’Action/le Réseau : – — –/Doctorants

Laboratoire/Entreprise : IMT Atlantique
Durée : 36 mois
Contact : seed-contact@imt-atlantique.fr
Date limite de publication : 2025-03-21

Contexte :
Domain and scientific/technical context

This project aims to develop computational imaging methods for low-field MRI [Arnold2023, Hennig2023]. Its aim is to develop low-cost, portable neuroimaging systems that integrate artificial intelligence (AI) [Iglesias2022] with low-field MRI technology. Unlike conventional MRI systems that rely on high magnetic fields (1.5-7T), this approach aims to democratise access to MRI by enabling imaging at the patient’s bedside.

The project is highly interdisciplinary, combining expertise in medical imaging, image processing, AI and neuroscience. It targets perinatal neuroimaging, in particular for premature newborns, for whom traditional MRI remains complex. By combining hardware development (in collaboration with the company Multiwave) and AI-driven image reconstruction, this project could redefine neuroimaging and improve its accessibility in clinical settings.

Scientific/technical challenges

The project tackles fundamental challenges in low-field MRI and computational imaging, necessitating a multidisciplinary approach. One of the very first challenges is related to signal-to-noise ratio limitations. The weak magnetic fields in low-field MRI produce inherently noisier signals, demanding innovative AI-driven denoising and reconstruction strategies tailored to low SNR conditions. A second challenge is related to the optimization of hardware design: developing a portable, cost-efficient MRI system requires a careful trade-off between coil design, acquisition protocols, and system portability while maintaining sufficient imaging resolution.

Our scientific objective will focus mainly on advanced AI methodologies. Incorporating physics-guided deep learning models that explicitly integrate the underlying MRI signal formation process to enhance reconstruction reliability and interpretability. To this end, part of the project will be dedicated to the development of efficient computational strategies: Achieving real-time image reconstruction necessitates optimized numerical solvers and meta-learning techniques for rapid inference at the point of care.

Sujet :

The project will leverage physics-informed deep learning for image reconstruction, integrating prior knowledge of MRI signal formation to enhance image quality. Variational optimization techniques [Fablet2021] will be explored to control the balance between acquired data and reconstructed images [Crockett2022], minimizing artifacts and improving clinical reliability. Meta-learning algorithms [Andrychowicz2016] will be implemented to optimize reconstruction efficiency for real-time bedside applications.

The expected results include the development of a fully functional image reconstruction prototype for low-field MRI, achieving millimetric resolution and demonstrating feasibility for neonatal brain imaging. The impact of the project extends beyond neonatal imaging, offering a scalable and accessible MRI solution for broader applications such as stroke detection [Yuen2022] and point-of-care diagnostics, particularly in low-resource settings. By bridging advances in AI and medical imaging, such a project has the potential to transform clinical neuroimaging and improve patient care worldwide.

Profil du candidat :
The skills required to carry out this work include machine learning, image processing and applied mathematics. Knowledge of computer science and programming (Python) will also be required in order to develop the associated algorithms.

Formation et compétences requises :
Master / Engineering school. Machine Learning, Deep Learning, Image Processing, Medical Imaging.

Adresse d’emploi :
IMT Atlantique, Campus de Brest.

The PhD student will stay 3 months each at an international academic and an industrial partners, respectively at University of Lausanne and Multiwave enterprise.

Document attaché : 202503101503_2025-SEED-image-reconstruction.pdf