Cascade-E-Yolov5s network for recognizing the ulcerative lesion subtypes in small intestinal.

Abstract

In endoscopy, accurately diagnosing small intestinal ulcers presents significant challenges due to the complex morphology, varying number, and extensive distribution of the lesions, which contribute to a reduced accuracy in immediate diagnosis. The definitive diagnosis typically relies on pathological analysis, laboratory investigations, and prolonged follow-up, often leading to diagnostic delays. This study introduces the Cascade-E-Yolov5s network, designed to improve the efficiency and accuracy of immediate ulcer diagnosis by intelligently identifying ulcer subtypes. The Cascade-E-Yolov5s network integrates EfficientNet for the classification of ulcer lesion images and SimAM-Yolov5s for detecting lesions on these classified images. In the SimAM-Yolov5s component, EfficientNet replaces the traditional backbone structure of Yolov5s, and enhancements such as the SIoU loss function and a simple, parameter-free attention module are incorporated to optimize model performance. The study utilized a dataset comprising 4909 ulcer images from 684 patients at Shanghai Changhai Hospital, encompassing four ulcer types: cryptogenic multifocal ulcerous stenosing enteritis, non-specific ulcer, small intestinal tuberculosis, and Crohn's disease. The experimental findings indicate that Cascade-E-Yolov5s surpasses conventional detection networks, achieving an average detection precision of 86.46% and a mean average precision at the IoU of 0.5 (mAP@0.5) of 82.20%. This model notably enhances the detection efficiency of small intestinal ulcer subtypes, thereby assisting clinicians in making more precise immediate diagnoses.