Video polyp segmentation with cross frame attention: CFVNet

Colorectal Cancer (CRC) is the second most lethal cancer and the third most common malignant tumor globally, accounting for approximately 10% of global cancer cases. Deep learning-based polyp segmentation methods have shown sufficient potential in improving the diagnostic efficiency. A large number of previous experiments were conducted on images, but compared with polyp videos, they lack attention to the correlation information between contexts. There are a large number of video frames in colonoscopy, which is extremely time-consuming and labor-intensive in clinical practice. Therefore, in order to increase the accuracy of the segmentation model, this paper proposes an excellent-performance video polyp segmentation framework named CFVNet (Cross Frame Video Network). This framework addresses the issues of unclear polyp localization and boundaries caused by factors such as lighting and angles, and captures the dynamic changes and temporal dependencies of polyps in the video. Our core idea is to make full use of the similar semantics between consecutive video frames, and incorporate a Video Context Segmentation module (VCS) and a Cross Frame Attention Module (CFA)) into the basic network structure, thereby improving the overall accuracy of segmentation. A large number of experiments conducted on the large-scale video polyp dataset SUN-SEG have demonstrated the superiority of our method.