摘要目的:尝试构建基于深度学习技术的胃肠道间质瘤（gastrointestinal stromal tumors，GISTs）与平滑肌瘤（leiomyomas，LM）超声内镜图像分类模型，并验证其鉴别诊断价值。方法:回顾性纳入2014年10月至2021年10月在苏州大学附属第二医院接受超声内镜检查且经外科手术或内镜下切除后病理确诊的69例GISTs和73例LM病例，每例病例选取1张清晰且有典型病变的超声内镜图片，利用留出法将每种疾病图片按训练集图片数比验证集图片数为8∶2的比例分入训练集和验证集，最终由113张（55张GISTs和58张LM）超声内镜图片组成训练集，由29张（14张GISTs和15张LM）超声内镜图片组成验证集，训练集用于对深度学习模型进行训练与优化，验证集用于对分类模型进行验证，主要观察指标包括鉴别诊断的灵敏度、特异度、阳性预测值、阴性预测值和准确率。结果:利用Resnet 34网络结构建立的分类模型对GISTs与LM进行鉴别诊断的准确率趋于0.89，较Resnet 50网络结构建立的分类模型（0.81）的分类性能更佳。基于Resnet 34网络结构构建的分类模型对验证集中超声内镜图片进行鉴别诊断的灵敏度、特异度、阳性预测值、阴性预测值和准确率分别为85.71%（12/14）（95% CI：67.38%~100.00%）、93.33%（14/15）（95% CI：80.71%~100.00%）、92.31%（12/13）（95% CI：77.82%~100.00%）、87.50%（14/16）（95% CI：71.30%~100.00%）和89.66%（26/29）（95% CI：78.57%~100.00%）。 结论:深度学习技术用于GISTs与LM超声内镜图像的鉴别诊断是可行的，可为临床医师对两者的鉴别提供辅助诊断意见。基于Resnet 34网络结构建立的深度学习模型对GISTs与LM超声内镜图像进行鉴别诊断的准确性较高。

abstractsObjective:To construct a classification model for endoscopic ultrasonography (EUS) images of gastrointestinal stromal tumors (GISTs) and leiomyomas (LM) based on deep learning technology, and to verify its value for differential diagnosis.Methods:From October 2014 to October 2021, 69 patients of GISTs and 73 of LM who underwent EUS and were pathologically confirmed by surgery or endoscopic resection in the Second Affiliated Hospital of Soochow University were retrospectively studied. One clear EUS image with typical lesion was selected for each case. Using the hold-out method, the images of each disease were divided into the training set and the validation set according to the ratio of the number of images in the training set to the number of images in the validation set, which was 8∶2. Finally, 113 EUS images (55 GISTs and 58 LM) were used to form the training set, and 29 EUS images (14 GISTs and 15 LM) were used to form the validation set. The training set was used to train and optimize the deep learning model, and the validation set was used to verify the classification model. The main observation indicators included the sensitivity, the specificity, the positive predictive value, the negative predictive value and the accuracy of differential diagnosis.Results:The accuracy of the classification model established by Resnet 34 network structure in the differential diagnosis of GISTs and LM tended to be 0.89, better than the classification model established by Resnet 50 network structure (0.81). The sensitivity, the specificity, the positive predictive value, the negative predictive value and the accuracy of the classification model based on Resnet 34 network structure for differentiating EUS images in the validation set were 85.71% (12/14, 95% CI: 67.38%-100.00%), 93.33% (14/15, 95% CI: 80.71%-100.00%), 92.31% (12/13, 95% CI: 77.82%-100.00%), 87.50% (14/16, 95% CI: 71.30%-100.00%) and 89.66% (26/29, 95% CI: 78.57%-100.00%), respectively. Conclusion:It is feasible to use deep learning technology in the differential diagnosis of EUS images of GISTs and LM, which can provide auxiliary diagnostic opinions for clinicians. The deep learning model based on Resnet 34 network structure shows higher accuracy in the differential diagnosis of EUS images of GISTs and LM.