CRPU-NET: a deep learning model based semantic segmentation for the detection of colorectal polyp in lower gastrointestinal tract

Abstract

Purpose. The objectives of the proposed work are twofold. Firstly, to develop a specialized light weight CRPU-Net for the segmentation of polyps in colonoscopy images. Secondly, to conduct a comparative analysis of the performance of CRPU-Net with implemented state-of-the-art models. Methods. We have utilized two distinct colonoscopy image datasets such as CVC-ColonDB and CVC-ClinicDB. This paper introduces the CRPU-Net, a novel approach for the automated segmentation of polyps in colorectal regions. A comprehensive series of experiments was conducted using the CRPU-Net, and its performance was compared with that of state-of-the-art models such as VGG16, VGG19, U-Net and ResUnet++. Additional analysis such as ablation study, generalizability test and 5-fold cross validation were performed. Results. The CRPU-Net achieved the segmentation accuracy of 96.42% compared to state-of-the-art model like ResUnet++ (90.91%). The Jaccard coefficient of 93.96% and Dice coefficient of 95.77% was obtained by comparing the segmentation performance of the CRPU-Net with ground truth. Conclusion. The CRPU-Net exhibits outstanding performance in Segmentation of polyp and holds promise for integration into colonoscopy devices enabling efficient operation.