基于深度迁移学习模型实现股骨头坏死与其他髋部疾病的X线片鉴别诊断

A deep transfer learning method using plain radiographs for the differential diagnosis of osteonecrosis of the femoral head with other hip diseases

目的采用深度迁移学习方法实现基于正位X线图像的股骨头坏死(osteonecrosis of femoral head,ONFH)、发育性髋关节发育不良(developmental dysplasia of hip,DDH)与其他常见髋关节疾病的鉴别诊断。方法回顾性收集2018年1月至2020年12月在广州中医药大学第一附属医院就诊的ONFH、DDH及原发性髋关节骨关节炎、非感染性炎性髋关节病、股骨颈骨折等髋关节疾病患者的髋关节正位X线图像,建立临床数据集。通过图像旋转、翻转形式的数据增强方法扩展数据集,然后将数据集随机平分为训练集和测试集。将可迁移归一化技术引入ResNet-152深层神经网络模型,替换原有的批归一化技术,建立新的深度迁移学习模型。通过训练集训练该模型,随后评价该模型在测试集基于髋关节正位X线图像实现人工智能区分ONFH、DDH及其他髋关节疾病的效果。结果临床数据集共纳入1024髋的正位X线图像,共计542髋ONFH、296髋DDH、186髋其他髋关节疾病(原发性骨关节炎56髋、非感染性炎性髋关节病85髋、股骨颈骨折45髋)。通过数据增强方法扩展为包含6144髋的数据集。在训练集上对深度迁移学习模型进行100050次训练。该模型区分ONFH及其他髋关节疾病的二分类准确率最佳值达95.80%,区分ONFH、DDH及其他髋关节疾病的三分类准确率最佳值达91.40%。模型重复训练至50000次后分类准确率达到平台期。二分类及三分类任务平台期的准确率平均为95.35%[95%CI(95.33%,95.37%)]和90.85%[95%CI(90.82%,90.87%)]。结论深度迁移学习模型在初诊环节,可基于简便、经济的正位X线图像区分ONFH、DDH与其他的常见髋关节疾病。

Objective To develop a deep transfer learning method for the differential diagnosis of osteonecrosis of the femoral head(ONFH)with other common hip diseases using anteroposterior hip radiographs.Methods Patients suffering from ONFH,DDH,and other hip diseases including primary hip osteoarthritis,non-infectious inflammatory hip disease,and femoral neck fracture treated in the First Affiliated Hospital of Guangzhou University of Chinese Medicine from January 2018 to December 2020 were enrolled in the study.A clinical data set containing anteroposterior hip radiographs of the eligible patients was created.Data augmentation by rotating and flipping images was performed to enlarge the data set,then the data set was divided equally into a training data set and a testing data set.The ResNet-152,a deep neural network model,was used in the study,but the original Batch Normalization was replaced with Transferable Normalization to construct a novel deep transfer learning model.The model was trained to distinguish ONFH and DDH from other common hip diseases using anteroposterior hip radiographs on the training data set and its classification performance was evaluated on the testing data set.Results The clinical data set was comprised of anteroposterior hip radiographs of 1024 hips,including 542 with ONFH,296 with DDH,and 186 with other common hip diseases(56 hips with primary osteoarthritis,85 hips with non-infectious inflammatory osteoarthritis,45 hips with femoral neck fracture).After data augmentation,the size of the data set multiplied to 6144.The model was trained 100050 times in each task.Accuracy was used as the representative parameter to evaluate the performance of the model.In the binary classification task to identify ONFH,the best accuracy was 95.80%.As for the multi-classification task for classification of ONFH and DDH from other hip diseases,the best accuracy was 91.40%.The plateau of the model was observed in each task after 50000 times of training.The mean accuracy in plateaus was 95.35%(95%CI:95.33%,95.37%),and 90.85%(95%CI:90.82%,90.87%),respectively.Conclusion The present study proves the encouraging performance of a deep transfer learning method for the first-visit classification of ONFH,DDH,and other hip diseases using the convenient and economical anteroposterior hip radiographs.