基于GE-YOLO的消化内镜下异常区域实时目标检测方法  

Real-Time Target Detection of Abnormal Regions in Gastrointestinal Endoscopy Based on GE-YOLO

作　　者：范姗慧 赖劲涛 韦尚光 魏凯华 范一宏[2] 吕宾[2] 厉力华 Fan Shanhui;Lai Jintao;Wei Shangguang;Wei Kaihua;Fan Yihong;Lv Bin;Li Lihua(School of Automation(School of Artificial Intelligence),Hangzhou Dianzi University,Hangzhou 310018,China;Department of Gastroenterology,Zhejiang Provincial Hospital of Traditional Chinese Medicine,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学自动化学院(人工智能学院),杭州310018 [2]浙江省中医院消化内科

出　　处：《中国生物医学工程学报》2024年第4期385-398,共14页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(62271182,81601530);温州市基础性公益科研项目(Y2023139);浙江省科技计划项目(公益技术应用研究)(2017C33143)。

摘　　要：消化内镜是临床常用的消化道检查手段,在消化道疾病的早期诊断和治疗中具有重要作用。但常规内镜检查需要由专业医生操作并实时观察视频以确定病灶点,极度依赖医生经验,主观性强且容易造成漏检和误检。本研究提出了一种基于改进YOLOv7-tiny的消化内镜下异常区域实时检测方法:GE-YOLO。该方法以YOLOv7-tiny为基础框架,使用两种不同的特征提取模块(C3模块和P-ELAN模块)构建骨干特征提取网络,提高网络的特征提取能力;使用坐标卷积(CoordConv)取代上采样中的普通卷积,进一步提高模型对病灶的定位精度;使用部分卷积(PConv)取代特征提取模块中的3×3普通卷积,在保证模型性能的同时减少了模型的计算量和参数量,提升了模型的检测速度;最后使用基于IoU与归一化Wasserstein距离的联合损失函数,提升模型对微小病灶的敏感度。该模型利用Kvasir-Capsule数据集中含标记的图像(共4172张)进行了训练和测试,其平均精确率、召回率和F1得分分别达到了94.2%、97.2%和0.957,检测速度为60帧/s,与YOLOv7-tiny相比,精确率、召回率和F1得分分别提升了2.8%、12.0%和0.075。实验结果表明,本研究提出的方法能实现高精度的消化道病灶实时检测,有望部署于临床内镜检查设备,在检查过程中为医生提供实时辅助,从而大大提高内镜检查效率,具有重要的学术价值和临床意义。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(4172 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.

关 键 词：GE-YOLO 实时目标检测 异常区域 消化内镜 YOLOv7-tiny 

分 类 号：R318[医药卫生—生物医学工程]