Deep learning for segmentation of colorectal carcinomas on endoscopic ultrasound.

Abstract

Bowel-preserving local resection of early rectal cancer is less successful if the tumor infiltrates the muscularis propria as opposed to submucosal infiltration only. Magnetic resonance imaging currently lacks the spatial resolution to provide a reliable estimation of the infiltration depth. Endoscopic ultrasound (EUS) has better resolution, but its interpretation is investigator dependent. We hypothesize that automated image segmentation of EUS could be a way to standardize EUS interpretation. EUS media and outcome data were collected prospectively. Based on 373 expert manual segmentations, a convolutional neural network was developed to perform segmentation of the submucosa, muscularis propria, and tumors. The mean surface distance (MSD), maximal distance between segmentations (Hausdorff distance; HDD), and overlap (Dice similarity index; DSI) were calculated. The median MSD and HDD values were 3.2 and 17.7 pixels for the tumor, 3.4 and 24.7 pixels for the submucosa, and 2.6 and 20.0 pixels for the muscularis propria, respectively. The median DSI values for the tumor, submucosa, and muscularis propria were 0.82, 0.57, and 0.59, respectively. These values reflect good agreement between manual and deep learning segmentation. This study found encouraging results of using automated analysis of EUS images of early rectal cancer, supporting further exploration in clinical practice.