A Review of Colorectal Polyp Segmentation Methods

Colorectal cancer is a malignant tumor with a persistently high incidence worldwide, and the majority of cases arise from the malignant transformation of polyps. However, polyps in colonoscopic images exhibit significant variations in morphology and scale, and the boundaries between polyps and surrounding tissues are often indistinct, which makes manual detection particularly challenging. On the one hand, the diagnostic process relies heavily on physicians’ experience, which may lead to missed diagnoses and misdiagnoses. On the other hand, the limited number of experienced clinicians makes it difficult to meet patients’ needs in a timely manner. This diagnostic uncertainty, combined with the scarcity of medical resources, highlights the critical value of AI-assisted diagnostic systems. Developing high-performance polyp segmentation algorithms that accurately locate and delineate polyp regions can improve diagnostic efficiency and provide effective decision support for clinicians. Recent advances in deep learning have significantly improved colorectal polyp segmentation. Regarding network architectures, models have evolved from early U-shaped designs to Transformer-based architectures that capture global context. More recently, Mamba-like networks and dual-branch architectures have also been proposed. In terms of research paradigms, this paper focuses on the application of methods such as multi-scale feature fusion, attention mechanisms, and diffusion models. In addition, this study summarizes commonly used datasets in the field of polyp segmentation and provides a detailed description of evaluation metrics closely related to segmentation performance. Finally, this paper systematically analyzes the limitations of current colorectal polyp segmentation methods in clinical practice. Based on this analysis, it discusses potential directions and trends for future algorithm development, providing a reference for further research.