Towards improving multiple sclerosis lesion segmentation performance with synthetic lesions

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil biomédical, à finalité spécialisée

Abstract

enMultiple sclerosis (MS) is the most prevalent neurological disease in the young adult population, affecting around 2.5 million people worldwide. It is a chronic neuroinflammatory condition that causes lesions in various regions of the central nervous system (CNS). Magnetic resonance imaging (MRI) is widely used for diagnosis, as it enables the visualization of MS lesions. Automatic segmentation of these lesions can provide neurologists with valuable insights into disease progression and support more accurate diagnosis. However, the limited availability of clinical data can decrease the performance of segmentation algorithms. Data augmentation, particularly the generation of synthetic MS lesions on MRI scans, has the potential to improve the performance of deep learning models, leading to more reliable diagnostic tools. In this master’s thesis, a non-deep learning method for synthetic lesion generation is presented and then evaluated with a deep learning segmentation algorithm, namely nnU-Net. Synthetic lesions generated by our pipeline were inserted into pathological MRI scans and added to the in-domain training set. The resulting model was then compared to a baseline model trained without synthetic data, by evaluating both on an in-domain test set and two out-of-domain datasets. A slight, non-significant performance improvement (+1.2% in DSC when using synthetic lesions in the training set compared to the baseline model) was observed on the in-domain test set. Results on the out-of-domain datasets were more variable: on the first out-of-domain dataset, a significant improvement was observed (+1.1% in DSC), while on the second one, a significant performance drop occurred (–4.4% in DSC). This suggests that the use of synthetic lesions alone is not sufficient to overcome the model’s limitations in terms of generalizability outside of domain. However, the increased segmentation accuracy in-domain suggests better generalization within the domain, which is encouraging for future work.