One-Shot Segmentation of Novel White Matter Tracts via Extensive Data Augmentation

Deep learning based methods have achieved state-of-the-art performance for automated white matter (WM) tract segmentation. In these methods, the segmentation model needs to be trained with a large number of manually annotated scans, which can accumulated throughout time. When novel WM tracts—i.e., tracts not included in existing tracts—are segmented, additional annotations need collected. Since annotation is time-consuming and costly, it desirable make only few training model, previous work has addressed this problem by transferring knowledge learned segmenting tracts. However, accurate still challenging one-shot setting, where one scan work, we explore setting data extremely scarce, on transfer framework, propose further perform extensive augmentation single scan, synthetic produced. We designed several different strategies that mask out regions augmentation. To avoid from potentially conflicting information produced strategies, choose each strategy separately network obtain multiple models. Then, results given models are ensembled final Our method was evaluated public in-house datasets. The experimental show our improves accuracy