DeepPoly: Deep Learning-Based Polyps Segmentation and Classification for Autonomous Colonoscopy Examination

Abstract

Colorectal cancer (CRC) is the third most common cause of cancer-related deaths in the United States and is anticipated to cause another 52,580 deaths in 2023. The standard medical procedure for screening and treating colorectal disease is a colonoscopy. By effectively examining the colonoscopy to identify precancerous polyps early and remove them before they become cancerous, CRC mortality can be lowered significantly. Manual colonoscopy examination for precancerous polyps detection is time-consuming, tedious, and prone to human error. Automatic segmentation and analysis could be fast and practical; however, existing automated methods fail to attain adequate accuracy in polyps segmentation. Moreover, these methods do not assess the risk of detected polyps. In this paper, we proposed an autonomous CRC screening method to detect polyps and assess their potential threats. The proposed method utilized DoubleU-Net for polyps segmentation and Vision Transformer (ViT) for classifying them based on their risks. The proposed method has achieved a mean dice-coefficient of 0.834 and 0.956 in segmentation for the Endotech challenge and Kvasir-SEG dataset, accordingly outperforming the existing state-of-the-art polyps segmentation. Then, this method classified the segmented polyps as hyper-plastic or adenomatous with 99% test accuracy.

Keywords

• Colorectal cancer
• colonoscopy
• polyps segmentation
• polyps classification
• EndoTect challenge
• Kvasir