Research on gastrointestinal polyp detection method based on improved YOLOv7.

Introduction: In the medical field, the detection of gastrointestinal polyps through endoscopic images is confronted with challenges such as complex backgrounds and numerous irrelevant factors. These issues often lead to a decrease in detection rates, an increase in missed detections, and a rise in the misdiagnosis rate of gastrointestinal lesions during early diagnosis. Methods: To enhance diagnostic accuracy, this paper proposes a gastrointestinal polyp detection method based on an improved YOLOv7 model. This method introduces the ECANet attention mechanism in both the head and neck of the YOLOv7 network structure to reduce the interference of image backgrounds and irrelevant factors, thereby improving the detection performance of the model. Furthermore, by replacing the loss function CIoU with EIoU, the improved model is able to better predict the bounding boxes of gastrointestinal polyps, getting closer to the real boxes, thereby enhancing the accuracy of model detection. Results: The models were compared on the Kvasir-SEG gastrointestinal polyp dataset. The precision of the improved model YOLOv7 (EIoU + ECANet) was 94%, the recall rate was 88.7%, and the mean average precision was 92.9%. Compared with the original model, all three indices have been improved. Discussion: The proposed YOLOv7 (EIoU + ECANet) model has strong robustness and generalization ability in the detection of gastrointestinal polyps.