Automatic Gland Segmentation for Detection of CRC Using Enhanced SegNet Neural Network

Abstract

Colorectal cancer (CRC) is a sort of cancer that begins in the rectum or large intestine, generally showing more impact on adults. It generally starts as tiny, non-cancerous (benign) lumps of cells called polyps that form on the colon's inward lining. Detection of colorectal cancers in early stages can reduce the risk of mortality rate. Detection of polyps through screening tests may fail because of the length of the rectum and the tiny size of the polyps. Computer-based biomedical imagining might help the radiologists sift through the information and analyse clinical pathology images better. The majority of research focuses on gland segmentation using CNN architectures, but accuracy is still challenging. In this work, encoder-decoder-based deep learning model was proposed for morphological colon gland segmentation. This research of enhanced SegNet architecture improved the system’s overall performance and, when tested on the standard benchmark Warwick-QU dataset for colonoscopy images. We trained the system by the training dataset consisting of 7225 images, and evaluation of the model is done using a test dataset divided into two parts: A (60 images), B (20 images). The results show that the proposed model is comparable to all the existing models regarding the same dataset and attain segmentation accuracy of 0.924 on Part A, 0.861 on Part B. The shape similarity of 83.145 on Part A, 88.642 on Part B, is considered adequate compared with the existing methods.