Self-supervised Multi-task Representation Learning for Sequential Medical Images

Self-supervised representation learning has achieved promising results for downstream visual tasks in natural images. However, its use the medical domain, where there is an underlying anatomical structural similarity, remains underexplored. To address this shortcoming, we propose a self-supervised multi-task framework sequential 2D images, which explicitly aims to exploit structures via multiple pretext tasks. Unlike current state-of-the-art methods, are designed only pre-train encoder instance discrimination tasks, proposed can and decoder at same time dense prediction We evaluate representations extracted by on two public whole heart segmentation datasets from different domains. The experimental show that our outperforms MoCo V2, strong baseline. Given small amount of labeled data, networks pre-trained unlabeled data achieve better than their counterparts trained standard supervised approaches.