Multi-classification of colorectal polyps with fused residual attention

Multi-classification of colorectal polyps using endoscopic images is crucial for enhancing clinical diagnostic accuracy and reducing colorectal cancer mortality. Accurately classifying colorectal polyps poses significant challenges due to blurred lesion boundaries, varying intra-class scales, and high inter-class similarities. To address these challenges, we propose the Fused Residual Attention Network (FRAN) for colorectal polyp classification. FRAN employs a dual-branch structure to emphasize both semantic and detailed information. The Residual Attention Learning mechanism enhances lesion region detection, while Global Dependent Self-Attention captures global context. Additionally, the Edge Feature Fusion module, combined with Semantic Alignment, mitigates semantic loss during upsampling and captures edge-detailed features. We evaluated FRA on a private four-class colorectal polyp dataset, the three-class public Kvasir dataset, the three-class public HyperKvasir dataset, and the four-class public PICCOLO dataset. The overall classification accuracies achieved are 85.73% and 97.16%, respectively, which are higher than those of the compared state-of-the-art colorectal polyp classification algorithms. Our approach effectively highlights critical regions and maintains detailed information, thereby offering a robust solution to the challenges in colorectal polyp classification.