Conflunet: Improving Confluent Lesion Identification In Multiple Sclerosis With Instance Segmentation

In multiple sclerosis (MS), white matter lesion (WML) instance masks are very relevant to enhance diagnosis and disease monitoring. Yet, all existing automated WML segmentation methods aim at improving a semantic segmentation model, and postprocessing it to group voxels together into lesion instances. A large majority of studies use connected components (CC) analysis for this final step. In this paper, we show both theoretically and empirically that CC is suboptimal for WML instance segmentation due to the presence of confluent lesions (CLs), i.e. lesions whose segmentation encompasses two or more individual lesions. We address this issue by proposing ConfLUNet - the first end-to-end instance segmentation model designed to detect and segment WML instances in MS. We evaluate ConfLUNet against two baseline methods, and show that it improves lesion detection metrics while maintaining similar segmentation performance. The results shown in this paper pave the way for more in-depth analysis of instance segmentation applied in the context of MS. Source code is available on Github.