Weakly-Supervised Gland Instance Segmentation based on Point Labeling

Colon cancer is a common digestive tract cancer occurring in the colon. How to effectively segment the glands in the pathological image of the colon is a challenging task. As well as distinguishing the glands from the background, it is necessary to distinguish the boundaries between the different instances of the glands. Common instance segmentation methods require pixel-level instance annotation, which takes a lot of time and resources. The pixel-level weakly supervised instance segmentation method cannot provide sufficient supervision, which makes the quality of gland segmentation poor. In this paper, we propose a weakly supervised gland segmentation method based on point labeling. We trained a glandular point detection model to predict high confidence points in gland images using the supervision information of point labeling. Then, the high-confidence point-assisted training instance segmentation model is used to implement the instance segmentation of the glands. We tested our method on the 2015 MICCAI Gland Challenge dataset, and the experimental results show that our method can effectively segment the instance of the Gland, and its performance is even better than that part of the method by training using the fully supervised.