Abstract

Info

Title:: Colorectal polyp segmentation in endoscopic images using DeepLab V3+

Author(s):: ZHU Shiqi1; 2; XU Chang1; 2; ZHOU Xin3; LIU Lu1; 2; LIN Jiaxi1; 2; YIN Minyue1; 2; LIU Xiaolin1; 2; XU Chunfang1; 2; ZHU Jinzhou1; 2; 1. Department of Gastroenterology, the First Affiliated Hospital of Soochow University, Suzhou 215006, China 2. Suzhou Clinical Centre of Digestive Disease, Suzhou 215006, China 3.Department of Gastroenterology, Jintan Affiliated Hospital of Jiangsu University, Changzhou 213200, China

Keywords:: Keywords: semantic segmentation deep learning DeepLab V3+ colorectal polyp

Abstract:: Abstract: Objective To establish a semantic segmentation model for colorectal polyps in endoscopic images based on DeepLab V3+. Methods A total of 1 500 endoscopic images of colorectal polyps were collected, including 1 000 from Hyper-Kvasir public dataset and 500 from the First Affiliated Hospital of Soochow University, and randomly divided into training set (n=1 200) and validation set (n=300). Meanwhile, the images from Jintan Affiliated Hospital of Jiangsu University were collected as test set (n=220). After the endoscopic image segmentation and labeling, the images and masks were loaded into a deep learning neural network with DeepLab V3+ as the architecture for training, thereby developing a semantic segmentation model. Results The accuracy, mean intersection over union, and Dice coefficient of the developed model were 97.2%, 85.8% and 0.924 in the internal validation set, and 98.0%, 80.1% and 0.890 in the external test set. Conclusion The DeepLab V3+ based semantic segmentation model for colorectal polyps in endoscopic images exhibits excellent performances, and it can serve as an effective method for the detection and diagnosis of colorectal polyps.