基于纹理特征的AutoML在NBI-ME判断食管癌分期中的应用

Application of AutoML Based on Texture Features in Judging the Staging of Esophageal Carcinoma by NBI-ME

目的探讨基于纹理特征的自动化机器学习(Automated Machine Learning,AutoML)在窄带成像技术结合放大内镜(Narrow-Band Imaging-Magnification Endoscopy,NBI-ME)图片中区分早期和进展期食管鳞癌的应用。方法收集苏州大学附属第一医院内镜中心食管鳞癌NBI-ME图片1507张,随机分为训练集(1264张)和验证集(243张)。使用MATLAB软件,提取整张内镜图片,共计32个纹理特征变量。将上述变量载入H2O平台进行AutoML二分类建模。另收集苏州大学附属第二医院内镜图片(278张)作为外部测试集。同时邀请1名低年资和1名高年资内镜医生对外部测试集进行判读。采用受试者工作特征(Receiver Operating Characteristic,ROC)曲线下面积(Area Under Curve,AUC)和准确度(Accuracy,ACC)等评估鉴别效能。结果基于RF算法的AutoML模型在外部测试集中表现最优,其AUC为0.975,ACC为0.939,显著优于其他模型,包括传统的GLM(AUC:0.776、ACC:0.687)和XGBoost模型(AUC:0.968、ACC:0.863);同时也优于低年资内镜医生(AUC:0.868、ACC:0.871)和高年资内镜医生(AUC:0.919、ACC:0.921)。结论基于内镜图片纹理特征的AutoML模型在食管早癌和进展期癌区别中展现出优秀的鉴别能力。

Objective To explore the application of automated machine learning(AutoML)based on the texture features in distinguishing early and advanced stages of esophageal squamous cell carcinoma by magnifying endoscopy with narrow-band imaging(NBI-ME).Methods A total of 1507 NBI-ME images of esophageal squamous cell carcinoma were collected from the Endoscopy Centre of The First Affiliated Hospital of Soochow University.These images were randomly divided into a training set(n=1264)and a validation set(n=243).MATLAB software was used to extract a total of 32 texture features from the whole endoscopic images.The above variables were loaded into the H2O platform for AutoML binary classification modeling.In addition,278 endoscopic images from the Second Affiliated Hospital of Soochow University were collected as an external test set.Moreover,one junior endoscopist and one senior endoscopist were invited to interpret the images from the external test set.Receiver operating characteristic(ROC)area under curve(AUC)and accuracy(ACC)were used to evaluate the efficiency of identification.Results The AutoML model based on random forest algorithm performed best in the external test set,with AUC of 0.975 and ACC of 0.939,which was significantly better than other models,including the generalized linear model(AUC:0.776,ACC:0.687)and the extreme gradient boosting model(AUC:0.968,ACC:0.863).It was also better than that of the junior endoscopist(AUC:0.868,ACC:0.871)and the senior endoscopist(AUC:0.919,ACC:0.921).Conclusion The AutoML model based on the texture features of endoscopic images shows excellent discriminative ability in judging the staging of esophageal carcinoma.