基于深度卷积神经网络的克罗恩病肛瘘磁共振成像诊断模型初探

A preliminary exploration of MRI diagnostic model of perianal fistulizing Crohn′s disease based on deep convolutional neural networks

目的初步探索基于深度卷积神经网络(DCNN)构建的克罗恩病(CD)肛瘘磁共振成像(MRI)诊断模型效能。方法采用回顾性研究方法,随机纳入2014年1月至2019年12月中山大学附属第六医院收治的200例初诊CD肛瘘患者和200例初诊腺源性肛瘘患者,每组按8∶1∶1分配至训练集、验证集和测试集。收集所有患者肛管MRI图像,预处理增强图像质量。采用Pytorch深度学习框架和Windows10计算机操作系统,基于4种DCNN(MobileNetV2、VGG11、ResNet18和ResNet34)构建CD肛瘘和腺源性肛瘘的MRI鉴别诊断模型。每种模型根据是否结合迁移学习策略,分为迁移学习型(T)和非迁移学习型(U)。首先,输入训练集(CD肛瘘和腺源性肛瘘患者各160例,共78321张MRI图像)图像数据,迭代训练至损失最小。然后,根据验证集(CD肛瘘和腺源性肛瘘患者各20例,共9697张MRI图像)的结果选择最佳的训练模型。最后,在测试集(CD肛瘘和腺源性肛瘘患者各20例,共9260张MRI图像)进行诊断效能评估。绘制每种预测模型的受试者操作特征(ROC)曲线并计算曲线下面积(AUC)。采用DeLong检验比较不同模型之间以及预测模型与不同年资放射科医生之间AUC的差异。结果结合迁移学习策略的4种诊断模型的效能分别为MobileNetV2-T(AUC=0.943,95%CI:0.820~0.991),VGG11-T(AUC=0.935,95%CI:0.810~0.988),ResNet18-T(AUC=0.920,95%CI:0.789~0.988),ResNet34-T(AUC=0.929,95%CI:0.801~0.986)。结合迁移学习策略的4种模型AUC均高于低年资放射科医生(均P<0.05),与高年资放射科医生的差异均无统计学意义(均P>0.05)。结论采用基于DCNN的深度学习技术,结合迁移学习策略和高分辨率肛管MRI构建CD肛瘘的病因诊断模型具有可行性。

Objective To evaluate the efficacy of magnetic resonance imaging(MRI)diagnostic model of perianal fistulizing Crohn′s disease(pfCD)based on deep convolutional neural networks(DCNN).Methods A restrospective study was conducted.The patients with pfCD of initial diagnosis(n=200)and the patients with cryptoglandular anal fistula(CAF)of initial diagnosis(n=200)were enrolled randomly in the Sixth Affiliated Hospital of Sun Yat-sen University from January 2014 to December 2019.The patients were assigned to the training,validation and test sets at a ratio of 8∶1∶1 in each group.The anal MRI images of all the patients were collected and preprocessed to enhance the quality of images.Using the Pytorch deep learning framework and Windows 10 computer operating system,the MRI diagnostic model of pfCD and CAF was constructed based on 4 DCNNs(MobileNetV2,VGG11,ResNet18 and ResNet34).Each model was divided into transfer learning(T)and untransfer learning(U)types based on whether it incorporated transfer learning strategy.First,the image data of training set(160 pfCD and 160 CAF patients,a total of 78321 MRI images)was input,and the training was iterated to minimize the loss.Then the best training model was selected based on the results of the validation set(20 pfCD and 20 CAF patients,a total of 9697 MRI images).Finally,diagnostic efficacy was evaluated on the test set(20 pfCD and 20 CAF patients,a total of 9260 MRI images).The receiver operating characteristic(ROC)curve for each model was drawn and the area under the curve(AUC)was calculated.The DeLong test was used to compare the difference in AUCs among different models and between models and radiologists with different seniorities.Results The efficacy of 4 models based on DCNN were MobileNetV2-T(AUC=0.943,95%CI:0.820-0.991),VGG11-T(AUC=0.935,95%CI:0.810-0.988),ResNet18-T(AUC=0.920,95%CI:0.789-0.988),ResNet34-T(AUC=0.929,95%CI:0.801-0.986),respectively.The AUCs of the 4 models combined with transfer learning strategy were higher than that of junior radiologist(all P<0.05),and there was no significant difference in AUCs between 4 models with transfer learning strategy and senior radiologist(all P>0.05).Conclusion The construction of diagnostic model of pfCD is feasible by using deep learning technology based on DCNN,transfer learning strategy and high-resolution anal MRI images.