White-light endoscopic colorectal lesion detection based on improved YOLOv7

Abstract

White-light endoscopy is an effective means of detecting colorectal lesions. Aiming at the phenomenon of misdetection and missed detection that are prone to occur in the detection process, this paper proposes an improved YOLOv7 model (CSS-YOLOv7) for lesion classification detection. First of all, the CNeB module is used to replace part of the ELAN module of the backbone and neck network, which can extract the features of the lesion area, reduce the number of network layers and model parameters, and improve the training speed. Second, colorectal lesion maps generated by white-light endoscopy have low resolution. Therefore, in response to this drawback, this study added the SPD-Conv module to the original model to effectively prevent fine-grained damage that may be caused by strided convolution. Finally, the SIOU loss function is used to enable the prediction frame to converge faster, and the precision and efficiency are effectively improved. The experimental results show that the recall rate of the model proposed in this paper is 92.3%, and mAP@.5 reaches 95.9%, which can promote the development of automatic detection of colorectal lesions.