Accurate identification of polyps in screening colonoscopies using convolutional neural networks

Abstract

Colorectal Cancer is one of the most prevalent cancers worldwide. One of the most critical factors for reducing the incidence of colorectal cancer is to increase the Adenoma Detection Rate (ADR), which is related to accurately detecting polyps during colonoscopy procedures. In this study, we developed a Convolutional Neural Network(CNN) architecture and utilised various techniques, including image processing and transfer learning. Our training dataset consisted of 1982 unique hand-labelled images extracted from 20 colonoscopy videos of different resolutions and 240 independent high-resolution colonoscopy images, all gathered by the Isfahan Gastroenterology and Hepatology Research Centre. Several experiments were conducted to assess the impact of CNN architectures on ADR. The experiments were designed to provide a fair evaluation of how the models would respond during a colonoscopy. The optimised CNN demonstrated excellent performance in polyp detection, achieving an Area Under the receiver operating characteristic Curve of 0.964 and an accuracy of 96.42. The results were pretty consistent regarding different video resolutions and types of polyps. In addition, their result was compared concerning three colonoscopy specialists who were presented with multiple images for a reduced amount of time to simulate routine procedures. The CNN outperformed the average accuracy of the specialists by 5. The proposed model demonstrates the potential to enhance and assist in the detection of adenomas and consequently contribute to higher prevention rates of colorectal cancer.