Method for automatic initialization of trainable active contours for instance segmentation in histological images

Abstract

The method of trainable active contour is one of the semi-automatic segmentation methods that can be applied to segment glands in histological images. In this paper, we propose a method for automatic initialization of trainable active contour model, which makes the segmentation method fully automatic. Using a U-Net like architecture, a preprocessed image segmentation masks is predicted for the input image, from which initial approximations of contours are calculated for each gland. The proposed method correctly marks 96.2 % part for the glands on the test set of the PATH-DT-MSU S1-v2 dataset. As a result, we get initial approximations located inside each gland in the image.