基于YOLOv8网络建立结直肠息肉实时检测的人工智能辅助系统(含视频)  

Establishment of an AI-assisted system(including video)for real-time detection of colorectal polyps based on the YOLOv8 network

作　　者：陈健[1,2] 孙斌 王甘红 夏开建 汤洪[5] 徐晓丹 周静洁 CHEN Jian;SUN Bin;WANG Ganhong;XIA Kaijian;TANG Hong;XU Xiaodan;ZHOU Jingjie(Department of Gastroenterology,Changshu Hospital Affiliated to Soochow University,Changshu 215500;Changshu Key Laboratory of Medical Artificial Intelligence and Big Data;Department of Gastroenterology,Changshu City Shanghu Central Hospital;Department of Gastroenterology,Changshu Traditional Chinese Medicine Hospital;Department of Gastroenterology,Xinzhuang People′s Hospital of Changshu City,China)

机构地区：[1]常熟市第一人民医院消化内科,江苏常熟215500 [2]常熟市医学人工智能与大数据重点实验室 [3]常熟市尚湖中心医院消化内科 [4]常熟市中医院消化内科 [5]常熟市辛庄人民医院消化内科

出　　处：《胃肠病学和肝病学杂志》2025年第4期538-545,共8页Chinese Journal of Gastroenterology and Hepatology

基　　金：常熟市医学人工智能与大数据重点实验室能力提升项目(CYZ202301);常熟市医药卫生科技计划项目(CS202019,CS202452);苏州卫生信息与健康医疗大数据学会项目(SZMIA2402)。

摘　　要：目的利用2023年新推出的YOLOv8m网络,开发一款人工智能辅助系统,旨在实现腺瘤性息肉的自动定位和诊断。方法使用4个结肠息肉数据集,总计包括9411张静态图像和25段视频。所涵盖的息肉类别包括增生性息肉和腺瘤性息肉。利用LabelMe工具对图像进行标注,并将标注数据转换成适用于深度学习模型训练的YOLO格式。在模型训练方面,采用预训练的YOLOv5m和YOLOv8m模型,并结合实时数据增强以及多种图像处理技术进行迁移学习训练。模型性能的评估采用多个指标,包括敏感性、特异性、假阳性率和检测速度(每秒帧数,frames per second,FPS)、平均精度(mean average precision,mAP)等。此外,还使用混淆矩阵进行详细评估,并将模型的性能与不同资历的医师进行比较分析。结果在对1411个息肉的验证集进行评估中,YOLOv8m模型在多项性能指标上超越了YOLOv5。YOLOv8m的整体准确率为98.58%,在腺瘤性息肉、增生性息肉检测的敏感性分别为98.06%和99.32%,特异性分别为99.33%和98.09%,不同类型息肉预测的mAP50为0.994。在与内镜医师的性能比较中,YOLOv8m模型在准确率(98.58%)和处理速度(60.61帧/s)方面均优于低年资(准确率为86.02%)和高年资内镜医师(准确率为93.14%),其处理速度是低年资内镜医师的67.2倍。结论基于YOLOv8m网络的深度学习模型能够快速、精确地检测与分类结直肠息肉,在辅助内镜医师提高腺瘤性息肉检出率方面展现出很大的应用潜力。Objective To develop an AI-assisted system using the newly introduced YOLOv8m network in 2023,designed to automate the localization and diagnosis of adenomatous polyps.Methods Four colon polyp datasets were utilized,totaling 9411 static images and 25 videos segments,covering categories included hyperplastic and adenomatous polyps.Images were annotated using LabelMe,with data converted into YOLO format suitable for deep learning model training.Pre-trained YOLOv5m and YOLOv8m models were employed for training,combined with real-time data augmentation and various image processing techniques.The model′s performance was assessed using multiple metrics,including sensitivity,specificity,false-positive rate,detection speed(frames per second,FPS),and mean average precision(mAP).A detailed evaluation was also conducted using confusion matrices,and the model′s performance was compared with physicians of different experience levels.Results In the validation set of 1411 polyps,the YOLOv8m model surpassed YOLOv5 in various performance metrics.YOLOv8m achieved an overall accuracy of 98.58%,with a sensitivity of 98.06%for adenomatous polyp detection and 99.32%for hyperplastic polyp detection.In terms of specificity,YOLOv8m performed at 99.33%for adenomatous polyps and 98.09%for hyperplastic polyps,with the mAP50 of 0.994 for different polyp types prediction.Compared to endoscopists,YOLOv8m excelled in accuracy(98.58%)and processing speed(60.61 frames/second),surpassing both junior(accuracy of 86.02%)and senior endoscopists(accuracy of 93.14%),processing at 67.2 times the speed of junior endoscopists.Conclusion The deep learning model based on the YOLOv8m network can rapidly and accurately detect and classify colorectal polyps,demonstrating significant potential in assisting endoscopists to enhance the adenoma detection rates.

关 键 词：结直肠息肉 深度学习 深度卷积神经网络 目标检测 YOLO 

分 类 号：R574.6[医药卫生—消化系统] TP181[医药卫生—内科学]