基于卷积神经网络胃食管结合部腺癌转移淋巴结人工智能诊断系统的建立

ESTABLISHMENT OF ARTIFICIAL INTELLIGENCE DIAGNOSTIC SYSTEM FOR METASTATIC LYMPH NODES IN GASTROESOPHAGEAL JUNCTION ADENOCARCINOMA BASED ON CONVOLUTIONAL NEURAL NETWORK

目的利用更快的区域卷积神经网络(Faster RCNN)构建胃食管结合部腺癌(AEG)转移淋巴结人工智能(AI)自动诊断系统,协助临床诊疗。方法回顾性选取2015年12月—2019年12月在青岛大学附属医院治疗的248例AEG病人,随机分为训练组和测试组。利用训练组CT图像训练Faster RCNN,建立AI诊断系统。对测试组病人的增强CT图像进行测试,并通过绘制受试者工作特征(ROC)曲线和计算ROC曲线下面积(AUC)、正确率、特异度、灵敏度、阳性预测值、阴性预测值、测试耗时等评估指标,分析AI诊断系统识别转移淋巴结的能力。结果AI系统诊断单张CT图片的时间约为0.15 s。AI系统诊断转移淋巴结的AUC为0.912,正确率为0.870,特异度为0.883,灵敏度为0.858,阳性预测值为0.892,阴性预测值为0.847。结论本研究建立的AI诊断系统识别AEG增强CT图像转移淋巴结的准确率高,且识别速度比临床医师快,具有辅助临床诊疗的潜力。

Objective To build an artificial intelligence(AI)-based automatic diagnostic system for metastatic lymph nodes in gastroesophageal junction adenocarcinoma(AEG)using a faster region-based convolutional neural network(Faster RCNN)to assist clinical diagnosis and treatment.Methods A total of 248 patients with AEG treated in The Affiliated Hospital of Qingdao University from December 2015 to December 2019 were retrospectively selected and randomly divided into training and test groups.The computed tomography(CT)images of the training group were used to train the Faster RCNN to establish an AI-based diagnostic system.The system was tested using the enhanced CT images of the test group.We analyzed the ability of the system to identify metastatic lymph nodes by drawing a receiver operating characteristic curve and calculating the area under the curve(AUC),accuracy,specificity,sensitivity,positive predictive value,negative predictive value,and test time.Results It took about 0.15 s for the AI system to diagnose a single CT image.The AI system for diagnosing metastatic lymph nodes had:AUC,0.912;accuracy,0.870;specificity,0.883;sensitivity,0.858;positive predictive value,0.892;and negative predictive value,0.847.Conclusion The AI-based diagnostic system in this study shows high accuracy in identifying metastatic lymph nodes in AEG on enhanced CT images,with faster recognition speed than clinicians.It has the potential to assist clinical diagnosis and treatment.