摘要
目的:探讨基于整张手腕部X线数字摄影(DR)影像深度学习特征的人工智能骨龄评估(AI-BAA)方法。方法:收集本院11858例0~18岁左手腕部骨龄DR图像数据,随机提取20%为验证集、80%为训练集,图像预处理后在resnet101基础上构建多模态信息融合的深度学习模型,优化算法以实现骨龄回归,并通过热力图实现数据模型可视化。另收集本院新近0~17岁1217例骨龄影像数据作为测试集,检验模型效能。采用平均绝对误差(MAE)和散点图评估模型骨龄预测的准确性。结果:模型骨龄预测值和儿科放射医师诊断结果的散点图呈一致性分布,MAE验证集0~18岁整体为(0.469±0.396)岁、男性为(0.453±0.396)岁、女性为(0.480±0.395)岁,测试集0~17岁整体为(0.459±0.371)岁,男性为(0.432±0.334)岁,女性为(0.511±0.429)岁。结论:基于整张手腕部DR影像高阶特征的AI-BAA模型提供了可靠的自动化骨龄检测方法。
Objective:To explore a method of artificial intelligence(AI)bone age assessment(BAA)based on deep feature learning for whole hand digital radiograph.Methods:Eleven thousand eight hundred and fifty-eight left hand digital radiographs of aged from 0 to18 years were collected from shanghai children’s hospital,20%of the total data were randomly extracted as validation set and 80%of the total data were used as training set after preprocessing of standardized image,data enhancement and normalization.On the basis of resnet101,the deep learning model of multimodal information fusion was constructed.The loss function combining L1 loss and L2 loss was selected and optimized by Adam algorithm to build up a bone age regression model.Then the grad-cam method was used to draw the thermodynamic diagram of the concerned area and realize the visualization of the data model.In addition,1217 left hand digital radiographs of aged from 0 to17 years were collected as test data to validate the accuracy and generalization of the model.The mean absolute error(MAE)and scatter plot model were used to evaluate the accuracy of bone age prediction.Results:Bone ages in different gender and age were all accurately assessed by the model in validation set and test set.MAE of bone age predicted by the model and pediatric radiologists was(0.469±0.396)years with(0.453±0.396)years for males and(0.480±0.395)years for females in validation set,and(0.459±0.371)years with(0.432±0.334)years for male and(0.511±0.429)years for female in test set respectively.The scattered plot of bone age evaluated by the model and doctor showed a consistent distribution.Conclusion:The AI BAA model based on higher-order imaging feature of the whole hand digital radiograph may provide a new reliable method of automatic bone age prediction.
作者
李婷婷
杨秀军
王乾
任旭华
兰钧
于广军
李嫔
李莉红
文颖
陈旭
LI Ting—ting;YANG Xiu—jun;WANG Qian(Department of Radiology,Shanghai Children's Hospital,Shanghai Jiao Tong University,Shanghai 200062,P.K.C.)
出处
《中国数字医学》
2019年第11期29-33,共5页
China Digital Medicine
基金
上海交通大学医工交叉重点项目（编号：YG2017ZD08）~~
关键词
人工智能
深度特征学习
骨龄
数字化放射摄影
artificial intelligence
deep feature learning
bone age
digital radiography
