Curriculum Self-supervised Learning for Weakly-supervised Histopathological Image Segmentation

Histopathological image segmentation is the basic upstream task of pathomics cancer research. However, acquiring pixel-level annotations is extremely difficult and time-consuming. Moreover, it is also expertise-dependent. Researchers pay a lot of attention on how to reduce the annotation effort. One feasible solution is self-supervised learning. In this paper, we proposed a novel self-supervised learning strategy for histopathological image segmentation, which called Curriculum Self-supervised Learning (CSSL). In CSSL, we set up several pretext tasks which are highly related to the characteristics of histopathological images for the CNN model. And then, these pretext tasks are trained in a curriculum manner (easy to hard). By combining CSSL with Class Activation Map (CAM), we can achieve outstanding performance on a weakly-supervised tissue segmentation task, with no need of pixel-level annotation. We achieved Frequency-weighted Intersection over Union (FIoU) score of 0.6886 and Mean Intersection over Union (MIoU) score of 0.6985 for semantic segmentation, which exceeded the common self-supervised learning methods. Experiments were conducted to prove the effectiveness of proposed CSSL.