Automatic tissue segmentation by deep learning: from colorectal polyps in colonoscopy to abdominal organs in CT exam

Automatic tissue segmentation is extremely helpful in medical imaging related work. In this paper, we attempted to train two existing deep neural networks, SegNet and DeepLab, to solve two clinical imaging problems relevant to this topic. One is to locate the colorectal polyps in the colonoscopy images, and the other is to delineate the lung in CT images from axial direction. In order to enhance the segmentation capability of the two networks, the reversed version of long short term memory (LSTM) network are integrated with them by parallel connection. The performance is evaluated by mean intersection-over-union (IOU). We found that introducing LSTM is beneficial to segmentation of polyps, but not that significant for delineating the lung. The relevant results are reported in this work.