YOLOv5-CE-DAFF: A polyp detection model based on CBAM-ECA attention mechanism and dropout-based adaptive feature fusion

Jiao Zhang,Qiming Zhang,Chen Qiao,Zhandong Mei,Kai Ren,Jian Chen

DOI: https://doi.org/10.21203/rs.3.rs-2897200/v1

2023-01-01

Abstract:Abstract Colorectal cancer often develops from polyps, and timely detection and removal of polyps can effectively reduce colorectal cancer mortality. However, traditional medical imaging methods have low accuracy and slow detection speed, which cannot meet the requirements of practical use. For the above reasons, we propose a real-time polyp detection model called YOLOv5-CE-DAFF, which is based on the YOLOv5 model and incorporates the CBAM-ECA attention mechanism and dropout-based adaptive feature fusion. Firstly, to pay more attention to polyp features, we incorporate the CBAM-ECA attention module into the local cross-layer fusion structure. Secondly, we design an adaptive feature fusion of the prediction feature layer based on dropout, which increases the number of high-quality prediction boxes. Thirdly, we improve the positive and negative sample matching strategy to accelerate model convergence and improve detection accuracy. Additionally, we combine Soft-NMS and WBF to replace the original NMS algorithm, which retains as many high-quality prediction boxes as possible. The model proposed in this paper is tested on the colonoscopy dataset provided by Huashan Hospital. Numerical experiments show that YOLOv5-CE-DAFF has high detection accuracy and a low missed detection rate for colon polyps. Compared with the original YOLOv5 model, the proposed model significantly improves recognition accuracy and reduces missed detection rate.