Real-Time Target Detection of Abnormal Regions in Gastrointestinal Endoscopy Based on GE-YOLO
Gastrointestinal endoscopy is a common clinical examination in early diagnosis and monitoring of gastrointestinal diseases.However,this examination needs to be operated by a professional doctor to identify lesions in real-time,it is extremely dependent on the doctor's experience which is subjective and easy to cause missed and/or false detection.In this study,GE-YOLO,a real-time detection method for abnormal object under digestive endoscopy based on improved YOLOv7-tiny,was proposed.Using YOLOv7-tiny as the basic framework,the backbone feature extraction network was constructed by using two different feature extraction modules(C3 module and P-ELAN module)to improve the feature extraction capability of the network;and then the coordinate convolution(CoordConv)was used to replace the normal convolution in the up-sampling,which made the model localize the lesion more accurately;furthermore,partial convolution(PConv)was applied to replace the 3×3 convolution in the feature extraction module,which not only guarantee the model detection performance,but also greatly reduced the computation cost and parameter number,and improved the model detection speed;finally,a joint loss function based on IoU and normalized Wasserstein distance was used to make the model more sensitive to small lesions.This model was trained and tested on the labeled images(4 172 in total)in Kvasir-Capsule dataset.The average precision,recall and F1-score of GE-YOLO was 94.2％,97.2％and 0.957,respectively,and the detection speed was 60 frames per second,which had an improvement of 2.8％in precision,12.0％in recall and 0.075 in F1-score compared with the results achieved by YOLOv7-tiny.The promising results demonstrated this proposed method can achieve high-precision real-time diagnosis of digestive tract lesions,and is expected to be deployed in clinical endoscopy equipment to provide real-time assistance for doctors during the examination to improve the diagnostic efficiency,which has momentous clinical value and research significance.
NETL
NSTL
万方数据
