VoxSeP: semi-positive voxels assist self-supervised 3D medical segmentation

Medical image segmentation enjoys the advantage of understanding 3D contexts, but 3D networks are prone to over-fitting due to the limited amount of annotated data. This paper investigates self-supervised pre-training, i.e. , making use of unlabeled medical data to initialize 3D segmentation networks. We build our system upon contrastive learning, where the dependence on positive and negative samples obstructs it from satisfying performance on medical image datasets with fewer samples. To alleviate this issue, we present a novel proxy task that takes advantage of the property of human body similarity in medical scans, and defines the sub-volumes from the same position of different cases as Semi-Positive samples. Pre-trained on a mixed dataset containing 1254 CT volumes, the proposed approach, VoxSeP , transfers well to 4 downstream datasets with 2 different backbone networks. On both fully supervised and semi-supervised fine-tuning, VoxSeP achieves favorable averaged improvements ( 2 % and 4 % ), which surpass several state-of-the-art counterparts.