Multi-task learning for gland segmentation

Morphology of glands is used by pathologist to evaluate the malignancy degree of adenocarcinomas which is a common type of cancer. Automatic analysis of histopathology images is important for a scalable and objective diagnosis, and segmentation of glands is a key step in this process. In this paper, we propose a method to accurately separate the gland instances from each other. We formulate the gland segmentation as a multi-task learning problem and generate the segmentation maps for the gland objects, contours and touching boundaries simultaneously. Our method uses the advantage of end-to-end learning and can be adapted to different base networks. To evaluate the proposed method, we use the benchmark “MICCAI 2015 Gland Segmentation in Colon Histology Images Challenge” dataset. On base networks DeepLabV3+ and U-Net, we show the success of the proposed multi-task model over single-task models. Comparisons with the reported results of the challenge and the results of other state-of-the-art studies support the advantages of our method.