Enrichment, Borrowing, And Mining: A Data-Driven Approach To Colonoscopic Lesion Classification

The automatic detection and diagnosis of colorectal lesions during colonoscopy are vital for the prevention and treatment of colorectal cancer. To date, numerous ROI-based neural network models have been proposed to classify colonoscopic lesions. However, they commonly encounter two main issues: 1) Their performance is limited by the dataset size and the diversity of lesion types included; 2) They tend to directly extract deep features from the original Region of Interests (ROIs) of the lesion, neglecting features of the lesion’s edge background as well as detailed texture and color information. To address these issues, we first propose a publicly available colonoscopic lesion classification dataset that includes all major categories of intestinal lesions. Secondly, we innovatively utilize additional colonoscopy video detection data with consistency between frames to enhance the lesion classification model by leveraging an online-target network. Simultaneously, to better mine the characteristics of lesions in data, we propose an ROI broaden strategy and a deep-fusion module, enabling the model to learn lesion-related information such as lesion edges, morphology, texture, and color. Experiments show that our method outperforms existing supervised and semi-supervised state-of-the-art methods, demonstrating the effectiveness of our approach. The GitHub link for our work is https://github.com/Zoe-TAN/EBM-polyp-cls.