深度学习骨龄智能评估系统在重庆三峡库区的初步验证  

Preliminary Verification of Deep Learning Bone Age Intelligent Assessment System in Chongqing Three Gorges Reservoir

作　　者：曾晨 孔俊沣 钟雯 刘嵩 曾文兵 乔虹[1] 杜文威 赵勇 ZENG Chen;KONG Jun-feng;ZHONG Wen;LIU Song;ZENG Wen-bing;QIAO Hong;DU Wen-wei;ZHAO Yong(Department of Endocrinology,the Second Affiliated Hospital of Harbin Medical University,Harbin 150000,Heilongjiang,China;Department of Radiology,Chongqing University Three Gorges Hospital,Chongqing 404100,China;Department of Child Health Care,Chongqing University Three Gorges Hospital,Chongqing 404100,China;Department of Radiology,Chongqing Kaizhou District People's Hospital,Chongqing 405400,China)

机构地区：[1]哈尔滨医科大学附属第二医院内分泌科,黑龙江哈尔滨150000 [2]重庆大学附属三峡医院放射科,重庆404100 [3]重庆大学附属三峡医院儿童保健科,重庆404100 [4]重庆市开州区人民医院放射科,重庆405400

出　　处：《医学信息》2022年第7期69-72,共4页Journal of Medical Information

基　　金：重庆市万州区科卫联合医学科研项目(编号:wzstc-kw2020031)。

摘　　要：目的探讨适用于重庆三峡库区临床应用的深度学习骨龄智能评估系统。方法纳入来自重庆三峡库区并于2020年6月-2021年7月检查的2500例儿童手腕骨骨龄片为研究数据集,其中2100例为训练集、200例为验证集,另200例为测试集,采用中华05 RUS-CHN法评估骨龄。3名放射学专家及1名儿童内分泌专家评估骨龄的均值为参考金标准。AI模型(深度学习骨龄智能评估系统)和对照组医师(医师A、医师B)独立阅片,分别记录骨龄评估的平均绝对误差(MAE)、耗时、绝对误差小于0.5岁及1.0岁样本所占比例。结果AI模型MAE为0.46岁[95%CI(0.36,0.56)],完成1份骨龄评估用时(1.31±0.82)s;2名对照组医师分别与AI模型评估的MAE比较,差异均无统计学意义(P>0.05),评估用时长于AI模型,差异有统计学意义(P<0.05);当误差范围在±1.0岁及±0.5岁以内,AI模型评估骨龄准确率为92.50%及75.50%,与2名对照组医师比较,差异均无统计学意义(P>0.05)。结论基于深度学习的儿童骨龄智能评估系统准确性高、耗时短,可用于重庆三峡库区儿童青少年骨龄的辅助评估。Objective To explore a deep learning bone age intelligent assessment system suitable for clinical application in the Chongqing Three Gorges Reservoir area.Methods A total of 2500 children’s wrist bone age films from the Chongqing Three Gorges Reservoir area and examined from June 2020 to July 2021 were included as the research data set,of which 2100 cases were the training set,200 cases were the validation set,and the other 200 were the test set,using the Zhonghua 05 RUS-CHN method to assess bone age.Three radiologists and one pediatric endocrinologist assessed the mean value of bone age as the reference gold standard.The AI model(deep learning bone age intelligent assessment system)and the control group doctors(doctor A,doctor B)read the pictures independently,and record the average absolute error(MAE),time-consuming,and the proportion of samples with absolute error less than 0.5 years old and 1.0 years old.Results The MAE of the AI model was 0.46 years old[95%CI(0.36,0.56)],and it took(1.31±0.82)s to complete a bone age assessment.There was no statistically significant difference in MAE between the two control doctors and the AI model(P>0.05).The evaluation time was longer than that of the AI model,and the difference was statistically significant(P<0.05).When the error range was within ±1.0 years and±0.5 years,the accuracy of AI model assessment of bone age was 92.50% and 75.50%,and there is no statistically significant difference compared with the two control doctors(P>0.05).Conclusion The intelligent assessment system for children’s bone age based on deep learning has high accuracy and short time-consuming,and can be used for auxiliary assessment of children and adolescents’bone age in the Chongqing Three Gorges Reservoir area.

关 键 词：骨龄测评 深度学习 儿童 中华05RUS-CHN法 三峡库区 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]