人工智能YOLO-V8算法构建早期胃癌图像检测系统的临床意义  

Research on the construction of image detection system for early gastric cancer by artificial intelligence YOLO-V8 algorithm

作　　者：金涛 姚镇东 毛伯能 蒋建中 陈燕春 JIN Tao;YAO Zhendong;MAO Boneng;JIANG Jianzhong;CHEN Yanchun(Yixing Clinical College,Medical College of Yangzhou University,Yixing 214200,China)

机构地区：[1]扬州大学医学院宜兴临床学院,江苏宜兴214200

出　　处：《临床肿瘤学杂志》2025年第3期261-267,共7页Chinese Clinical Oncology

基　　金：宜兴市卫生健康委员会重大项目(YXKY202209)。

摘　　要：目的验证YOLO-V8在胃癌内镜图像上的泛化性能,以评估其准确性和稳定性,完善YOLO-V8系统,优化构建的EGC-YOLO-V8人工智能系统,完善参数配置。方法收集2020年至2022年宜兴市人民医院的胃镜图像病例,根据病理类型将其分为两类,包括早期胃癌图像(阳性样本)和非胃癌图像(阴性样本)。将数据集划分为训练集和测试集。建立训练集,调整模型的参数,利用测试集进行模型选择,测试分析,设定阈值(Threshold),记录敏感度(Sensitivity),特异度(Specificity),精确度(Precision),约登指数(Youden Index),使用受试者工作特征(Receiver Operating Characteristic,ROC)曲线评估方法,计算曲线下面积(Area Under the Curve,AUC),来取得最佳泛化效果时的参数设置。结果训练集1的AUC平均值0.769,对应最大约登指数0.614,敏感度0.723,特异度0.891,精确度0.844,对应阈值0.01;训练集2的AUC平均值0.804,对应最大约登指数0.669,敏感度0.688,特异度0.981,精确度0.882,对应阈值0.02。结论本研究最佳的阈值设定为0.02,最佳的训练集阳性和阴性样本的数量比例为1∶2,系统能体现出更好的泛化性能。Objective To verify the generalization performance of YOLO-V8 on the endoscopic images of gastric cancer,to evaluate its accuracy and stability,improve the YOLO-V8 system,optimize the EGC-YOLO-V8 AI system,and improve the parameters configuration.Methods Gastroscopic image cases from Yixing People′s Hospital from 2020 to 2022 were collected and divided into two categories according to their pathological types,including early gastric cancer images(positive samples)and non-gastric cancer images(negative samples);Divide the datasets into training and test sets.Establish the training sets,adjust the parameters of the model,select the model by using the test sets,test and analyze,set the threshold,record the sensitivity,specificity,precision,Youden index,using the Receiver Operating Characteristic(ROC)Curve assessment method,calculate the Area Under the Curve(AUC),to set the parameters for the best generalization effect.Results The average AUC of training set 1 is 0.769,corresponding to the maximum Youden index 0.614,sensitivity 0.723,specificity 0.891,precision 0.844,corresponding threshold 0.01.The average AUC of training set 2 is 0.804,corresponding to the the maximum Youden index 0.669,sensitivity 0.688,specificity 0.981,precision 0.882,corresponding threshold 0.02.Conclusion In this study,the best threshold is set at 0.02,and the ratio of positive and negative samples in the best training set is 1:2,the system could reflect better generalization performance.

关 键 词：早期胃癌 人工智能 YOLO-V8 内镜诊断 

分 类 号：R735.2[医药卫生—肿瘤]