A Histo-Puzzle Network for Weakly Supervised Semantic Segmentation of Histological Tissue Type

Abstract

Digital pathological images with a large range of Histological Tissue Types (HTTs) contain more sophisticated contours than natural images. In recent years, deep learning algorithms have been widely applied to assist HTT analysis in a weakly-supervised manner by exploiting the class activation maps (CAM). However, the previous methods tend to confusedly activate the most discriminative regions of feature maps, resulting in incomplete segmented contour. This paper proposes a Histo-Puzzle network to improve the HTTs classification and segmentation based on patch-level self-supervised learning. Specifically, our model separates the HTT images into tiled patches by a puzzle module. Then we train a classifier on the supervision of reconstructed CAMs and image-level labels simultaneously. Experiments are conducted on the digital pathology database with 51 hierarchical HTTs. The experimental results show that our proposed method outperforms previous state-of-the-art methods on segmentation tasks of morphological and functional types.