人工智能骨龄系统提高西藏放射科医师骨龄判读的准确性  

Artificial intelligence bone age system improves the accuracy of bone age interpretation by tibetan radiologists

作　　者：拉巴顿珠 次旦旺久[1] 边巴次仁 王凤丹[2] 陈适[3] 李正 吴红[5] 次旦旺姆[6] 潘慧[3] 银武[1] 金征宇[2] LABA Dunzhu;CIDAN Wangjiu;BIANBA Ciren;WANG Fengdan;CHEN Shi;LI Zheng;WU Hong;CIDAN Wangmu;PAN Hui;YIN Wu;JIN Zhengyu(Department of Radiology,People's Hospital of the Tibet Autonomous Region,Lhasa,Tibet 850000,China;Department of Radiology,Peking Union Medical College,Peking Union Medical College Hospital,Chinese Academy of Medical Sciences,Beijing 100730,China;Endocrinology Department of Peking Union Medical College,Peking Union Medical College Hospital,Chinese Academy of Medical Sciences,Beijing 100730,China;Hangzhou Yitu Medical Technology Co.,Ltd.Hangzhou 310000,China;Department of Paediatrics,People's Hospital of the Tibet Autonomous Region,Lhasa 850000,China;Department of Child Care,Tibet Autonomous Region People's Hospital,Lhasa 850000,China)

机构地区：[1]西藏自治区人民医院放射科,西藏拉萨850000 [2]中国医学科学院北京协和医学院北京协和医院放射科,北京100730 [3]中国医学科学院北京协和医学院北京协和医院内分泌科,北京100730 [4]杭州依图医疗科技有限公司,浙江杭州310000 [5]西藏自治区人民医院儿科,西藏拉萨850000 [6]西藏自治区人民医院儿保科,西藏拉萨850000

出　　处：《医学影像学杂志》2023年第11期2061-2065,共5页Journal of Medical Imaging

基　　金：国家自然科学基金青年科学基金项目(编号:82001900);西藏自治区自然科学基金组团式医学援藏项目[编号:XZ2020ZR-ZY01(Z)];北京协和医院中央高水平医院临床科研项目(编号:2022-PUMCH-A-003)。

摘　　要：目的探讨在人工智能(artificial intelligence,AI)骨龄系统辅助下评估骨龄的准确性。方法选取4~18岁儿童的左手腕部X线摄影片作为骨龄片,以北京协和医院2位专家依据Greulich-Pyle(GP)图谱法共同阅片的骨龄作为参考标准。2位来自西藏自治区人民医院的藏族放射科医师(R1和R2),分别在有无AI骨龄系统辅助的情况下进行GP图谱法骨龄判读。采用平均绝对偏差(mean absolute deviation,MAD)、Bland-Altman图和配对t检验对AI辅助前后的骨龄判读结果进行统计分析。结果共纳入390例儿童,其中藏族305例,汉族85例,平均年龄(11.6±4.9)岁。R1和R2判读骨龄的准确率由辅助前的90.3%和90.8%提高到AI辅助后的98.7%和97.4%,MAD由0.46岁和0.50岁降至AI辅助后的0.19岁和0.26岁,Bland-Altman图95%置信区间由1.35~1.58岁和1.15~1.40岁缩小至AI辅助后的0.61~0.73岁和0.87~1.02岁。R1和R2的骨龄判读结果在AI辅助前后差异有统计学意义(P≤0.001)。结论AI骨龄系统能提高判读骨龄的准确性,接近骨龄判读经验丰富的专家水平。Objective To evaluate whether the accuracy of Tibetan Radiologists in bone age(BA)interpretation could be improved with the assistance of a previously published artificial intelligence(AI)BA system.Methods Two Tibetan radiologists(R1 and R2)independently assessed left hand radiographs of children aged from 4 to 18 years who presented with trauma to Tibet Autonomous Region People’s Hospital between September 2013 and November 2019,using Greulich-Pyle(GP)atlas with and without AI assistance.Reference standards were taken as the interpretation results of two experienced reviewers from Peaking Union Medical College Hospital in a consensus manner based on the GP atlas.For statistical analysis,mean absolute deviation(MAD),Bland-Altman and paired t test were used to compare the BA results with and without AI assistance.Results There were 390 children(305 Tibetan and 85 Han),with a mean age of(11.6±4.9)years.The accuracy of R1 and R2 in assessing BA was improved from 90.3%and 90.8%without AI assistance to 98.7%to 97.4%after AI assistance;the MAD of R1 and R2 decreased from 0.46 years and 0.50 years to 0.19 years and 0.26 years after AI assistance,respectively.According the Bland-Altman map,95%confidence interval of R1 and R2 shrank 1.35~1.58 years and 1.15~1.40 years without AI to 0.61~0.73 years and 0.87~1.02 years with AI,respectively.The BA results of R1 and R2 with and without AI had statistically significant difference(P≤0.001).Conclusion AI improves Tibetan radiologists’performance of BA assessment in comparable to experienced experts.

关 键 词：骨龄 人工智能 X线摄影 

分 类 号：R726.8[医药卫生—儿科] R445.4[医药卫生—临床医学]