基于MRI单张图像的深度卷积神经网络对膝关节前交叉韧带撕裂的诊断价值

VALUE OF DEEP CONVOLUTIONAL NEURAL NETWORK BASED ON SINGLE MRI IMAGES IN DIAGNOSIS OF ANTERIOR CRUCIATE LIGAMENT TEARS OF THE KNEE

目的建立基于膝关节MRI单张图像的深度卷积神经网络(DCNN)模型,并分析其诊断前交叉韧带(ACL)撕裂的价值。方法收集2017年1月1日—2022年6月30日海军第971医院GreatPACS影像存档与通信系统中1663例受检者的膝关节MRI图像,经一名骨科专科医生在每例患者MRI图像中手动选取1张图像并进行ACL正常或撕裂(正常1111张,撕裂552张)标注。将所有图像按照83%和17%的比例随机分配到训练集(1383张)和测试集(280张)中,用以训练并测试搭建的ACL智能诊断DCNN模型。通过阳性预测值(PPV)、阴性预测值(NPV)、准确率、灵敏度、特异度、受试者工作特征曲线下面积(AUC)等指标评估模型性能。结果本研究成功设计并搭建了ACL智能诊断DCNN模型。该模型诊断ACL撕裂的PPV、NPV、准确率、灵敏度和特异度分别为52.99%、88.96%、73.93%、77.50%和72.50%,AUC值为0.602。结论基于MRI单张图像DCNN模型对于临床医生诊断ACL撕裂具有一定的辅助作用。

Objective To establish a deep convolutional neural network(DCNN)model based on the single MRI image of the knee,and to investigate its value in the diagnosis of anterior cruciate ligament(ACL)tears.Methods Knee MRI images were collected from 1663 subjects from the GreatPACS image archiving and communication system in No.971 Hospital of People’s Liberation Army Navy from January 1,2017 to June 30,2022,and one image was selected from the MRI images of each patient and was annotated as normal ACL or ACL tears by an orthopedic specialist,which obtained 1111 images with normal ACL and 552 images with ACL tears.The images were randomly assigned to the training set(1383 images)and the test set(280 images)at a ratio of 83%and 17%,respectively,to train and test the DCNN model established for the intelligent diagnosis of ACL.The performance of the model was evaluated by positive predictive value(PPV),negative predictive value(NPV),accuracy,sensitivity,specificity,and area under the ROC curve(AUC).Results A DCNN model was successfully established for the intelligent diagnosis of ACL.The test results of this model showed a PPV of 52.99%,an NPV of 88.96%,an accuracy of 73.93%,a sensitivity of 77.50%,and a specificity of 72.50%,with an AUC of 0.602.Conclusion The DCNN model based on single MRI images can help clinicians with the diagnosis of ACL tears.