Séminaire I&M, par Marc-Adrien Hostin

Marc-Adrien Hostin

Titre : Characterization of Neuromuscular Diseases: Biomarker evaluation, Deep Learning Segmentation and Synthesis of lower limb MRI 

Résumé :

Neuromuscular Disease (NMD) leads to muscle weakness and even loss of motor function. Intramuscular fat fraction, a key biomarker of NMD progression, can be measured by quantitative Magnetic Resonance Imaging (qMRI). The extraction of qMRI biomarkers requires muscle segmentation. We compared several segmentation methods using Deep Learning (DL), to replace the time-consuming step of manual segmentation. The nnUNet model stood out thanks to its automatic optimization of hyperparameters, and the robustness of the NMD biomarkers derived from its segmentations. From the segmented muscle areas, we used a radiomics-based approach to identify potential biomarkers of NMD progression. Our results showed that a set of texture descriptors could provide a fine-grained characterization of NMD progression, allowing to observe the evolution of the texture of pathological muscle tissue.  As NMDs are rare, DL model training suffers from a lack of data. We proposed to fill this gap by creating synthetic images of patients. Our model, ConText-GAN, combines DL and radiomics to generate images of patients by specifying the texture of their muscle tissue, a manifestation of NMD involvement. The synthetic images were evaluated as realistic and diverse, and their use enabled us to improve the performance of segmentation models.