DCNN-based Polyps Segmentation using Colonoscopy images

Abstract

Colorectal polyps, which are associated with colorectal cancer, can be detected using a colonoscopy. Using the data from colonoscopy images to segment polyps is crucial in medical practice because it provides critical data for identification and surgery. However, precise segmentation of polyps is difficult due to the following factors: the polyp-bordering mucosa boundary is not sharp, and polyps of the same type differ in texture, size, and color. We propose to use the DeepLabV3+ architecture for image segmentation for medical purposes by examining its segmentation results on colonoscopy images from the datasets Kvasir and CVC-ClinicDB. DeepLabV3+ generates an F1-score of 0.865 for CVC-ClinicDB on an NVIDIA A100 class Cloud-Based GPU. The model is divided into the following parts: an encoder that performs separable convolution on the input map and a decoder that up-samples the data provided by the encoder using transpose convolution. Our approach significantly enhances segmentation accuracy and offers a number of benefits with respect to generality and real-time segmentation efficiency, according to evaluations done quantitatively and qualitatively on the two datasets.