基于声像图深度残差网络ResNet模型自动诊断肾囊肿

Automatic diagnosis of renal cyst with depth residual network ResNet models based on ultrasonograms

目的分析基于声像图的深度残差网络ResNet模型自动诊断肾囊肿的效能。方法收集3670例患者的肾脏超声资料,其中2024例超声诊断肾囊肿、1646例肾脏正常,每例选取2幅肾脏声像图,共以7340幅肾脏声像图构建数据集;以其中6294幅(3238幅肾囊肿、3056幅正常肾)为训练集,1046幅(810幅囊肿、236幅正常肾)为测试集,分别采用梯度方向直方图(HOG)+支持向量机(SVM)方法及3种深度残差网络ResNet模型(ResNet18、ResNet34、ResNet50)进行诊断。以超声诊断结果为金标准,计算并比较4种方法诊断测试集肾囊肿的敏感度、特异度及准确率,并绘制受试者工作特征(ROC)曲线,获得曲线下面积(AUC)。结果ResNet34、ResNet50模型诊断测试集肾囊肿的敏感度、特异度及准确率均高于HOG+SVM方法及ResNet18模型(P均<0.01),且ResNet50模型的特异度和准确率均高于ResNet34模型(P均<0.05)。ROC曲线显示,HOG+SVM方法及ResNet18、ResNet34、ResNet50模型自动诊断肾囊肿的AUC分别为0.731[95%CI(0.691,0.771)]、0.754[95%CI(0.715,0.792)]、0.851[95%CI(0.819,0.884)]及0.892[95%CI(0.865,0.920)]。结论基于声像图的深度残差网络ResNet模型可自动诊断肾囊肿,以ResNet50模型效果最佳。

Objective To observe the efficacy of depth residual network ResNet models in automatic diagnosis of renal cyst based on ultrasonograms.Methods Data of renal ultrasound of 3670 patients were collected.Ultrasound physicians diagnosed renal cysts in 2024 cases,and the remaining 1646 cases were found with normal kidneys.Two images were collected for each case,and a dataset containing 7340 images was conducted,among which 6294 images including 3238 of renal cysts and 3056 of normal kidneys were divided into the training set,while other 1046 images,including 810 of renal cysts and 236 of normal kidneys were divided into the test set.The method of histogram of oriented gradients(HOG)+support vector machines(SVM)and three deep residual network ResNet models(ResNet18,ResNet34 and ResNet50)were used to analyze renal ultrasonograms.The diagnostic results of ultrasound physicians were taken as the gold standards,and the sensitivity,specificity and accuracy of the above 4 methods for diagnosing renal cysts in test set were calculated and compared.Receiver operating characteristic(ROC)curve was used to obtain the area under the curve(AUC).Results In test set,the sensitivity,specificity and accuracy of ResNet34 and ResNet50 models for diagnosing renal cysts were all higher than those of HOG+SVM method and ResNet18 model(all P<0.01),while the specificity and accuracy of ResNet50 model were higher than those of ResNet34 model(both P<0.05).ROC curve showed that the AUC of HOG+SVM and ResNet18,ResNet34 and ResNet50 models for automatic diagnosis of renal cyst was 0.731(95%CI[0.691,0.771]),0.754(95%CI[0.715,0.792]),0.851(95%CI[0.819,0.884])and 0.892(95%CI[0.865,0.920]),respectively.Conclusion Based on ultrasonograms,ResNet models deep residual network could be used for automatic diagnosis of renal cyst,and ResNet50 model had the best efficacy.