COPN-NET: DEEP LEARNING BASED COLORECTAL POLYP CLASSIFICATION VIA MOBILE REGNET AND DEEP BELIEF NETWORK

Abstract

Colorectal polyp classification (CPC) is essential for the early diagnosis and prevention of colorectal cancer (CC). However, existing methods often struggle to accurately classify polyps due to variations in size, shape, texture, illuminations and imaging artifacts in colonoscopy images. To address these challenges, this paper proposes COPN-NET, a novel deep learning framework combining Mobile RegNet with a Deep Belief Network (DBN). Adaptive Wiener Star filtering is first applied for noise suppression and contrast enhancement. Multi-scale feature extraction is performed using an optimized Mobile RegNet architecture, followed by classification using DBN to distinguish hyperplastic polyps, adenomas, and normal tissues. Experimental evaluations on public colonoscopy datasets demonstrate an accuracy of 95.8% and sensitivity of 96.1%.