Accurate and robust automated detection and segmentation of Multiple Sclerosis lesions in spinal cord MRI

Research Motivation. Multiple Sclerosis (MS) is a common and potentially debilitating disease, characterised by the occurrence of lesions in the central nervous system. Recently, the number of first-line disease-modifying treatments for MS has augmented significantly and highly effective second-line immunosuppressive treatments have become available (McGinley, Goldschmidt, and Rae-Grant 2021). With the increasing treatment options, each of them with different potential adverse events and efficacy, accurate patient assessment and close monitoring become paramount.

Magnetic Resonance Imaging (MRI) plays a central role in MS patients assessment and follow-up, allowing for the identification of MS lesions at an early stage, as well as for the detection of new MS lesions appearing over time. Notably, MS lesions in the spinal cord or brain stem are known to be prognostic of quick acquisition of disability, for which more aggressive treatment might be initiated. Also, the appearance of new lesions in the brain or spinal cord indicates that a treatment change has to be considered.

Nevertheless, the identification of MS lesions is a challenging and laborious task for radiologists, which most often results in missed lesions. This motivates the need for automated tools for MS lesion detection and segmentation that can provide an aid to clinicians in their studies and daily practice, making patient assessment more accurate, consistent and repeatable.

Challenges and Objectives. Deep Learning (DL)-based solutions have been successfully applied to automate the detection and segmentation of MS lesions in brain MRI (Commowick et al. 2021), and it has been shown that such tools can bring substantial aid in clinical practice (Combès et al. 2021). Despite the high diagnostic and prognostic value of spinal cord MRI, only a few attempts have been made to address the automated identification of MS lesions in this area (Gros et al. 2019). Indeed, specific challenges exist, namely the limited amount of annotated data and the large heterogeneity of spinal cord MRI acquisition protocols.

This project aims at developing an accurate and robust solution to the automated detection and segmentation of MS lesions in spinal cord MRI. With this aim, i) the labelled data and the large amount of unlabelled data available to us from the different institutions¹ will be leveraged to provide unique insights into the variability of MS spinal cord MRI data, and ii) a DL-based solution will be developed that builds upon existing approaches that seek to address the shifts across data distributions i.e. domain adaptation and generalisation (Ackaouy et al. 2020; He et al. 2021). Last, iii) the clinical significance of the project will be assessed, in terms of facilitating and robustifying clinical studies, and optimising the follow-up of MS patients in clinical practice.

Candidate Profile. We are seeking a highly motivated candidate with a background in image processing, machine learning and data analysis. The candidate should be familiar with deep learning techniques and have good programming skills (Python). Knowledge in medical imaging will be appreciated. Good communication skills are fundamental.

¹Empenn takes part in the OFSEP project, Observatoire Français de la Sclérose En Plaques, and collaborates with the neurology department of the Rennes CHU.