不同人工智能模型对基于手腕部DR影像的骨龄预测比较

Comparison of two Different Artificial Intelligence Models of Bone Age Detection Based on Hand DR Data

目的探讨2种深度学习模型对基于手腕部DR影像骨龄预测价值,为骨龄评估提供合适的人工智能模型。方法搜集本院11858例0~18岁骨龄检测的左手腕部DR影像资料,构建基于其影像传统关注局部区域(AIM1)或数据驱动整体区域(AIM2)深度学习特征的骨龄预测模型。应用2种模型分别对本院2018年2月1216例(男415例,女801例)8个月~17岁儿童骨龄左手腕部DR影像资料进行测试,比较其骨龄预测值及与医师读数平均绝对误差(MAE)的差异,并评估其相关性,P<0.05为差异有统计学意义。结果参照儿科放射科医师基于GP图谱的骨龄读数,准确率AIM1为90.87%,低于AIM2的94.73%(P=0.001);MAE值AIM1为0.441±0.434,高于AIM2的0.437±0.328(P=0.929);其中,AIM2对女孩骨龄预测值更接近医师骨龄读数(P=0.78),AIM1对男孩骨龄预测值更接近医师骨龄读数(P=0.914);骨龄预测值2种模型之间及其与医师骨龄读数均具有显著相关性(P<0.01)。结论基于整体手腕部DR影像数据驱动人工智能模型对骨龄预测准确性高于基于临床先验知识的人工智能模型。

Objective To explore the different values of two deep learning models in bone age prediction based on hand DR images,and to provide a suitable artificial intelligence model for bone age assessment.Methods Firstly,the DR data of the left hand of 11858 patients aged from birth to 18 years were collected to establish a prediction model of bone age based on the traditional ROI(AIM1)or whole(AIM2)deep learning characteristics of the DR images.The DR imaging data of left hand of 1216 children(415 males and 801 females)aged from 8 months to 17 years old were tested by two models.Comparing the predicted value of bone age with the mean absolute error(MAE)of doctor’s readings,and evaluating its correlation,P<0.05 that there was statistical significance.Results The accuracy of AIM1 was 90.87%,which was lower than 94.73%of AIM2(P=0.001)according to the GP-based bone age readings of pediatric radiologists.The MAE value of AIM1 was 0.441±0.434,which was higher than that of AIM2 0.437±0.328(P=0.929);Among them,AIM2 was closer to the doctor’s readings for girls(P=0.78),AIM1 was closer to the doctor’s readings for boys(P=0.914),and there was significant correlation between the two models and doctor’s readings(P<0.01).Conclusion Data-driven artificial intelligence model based on global hand DR image is more accurate in predicting bone age than artificial intelligence model based on prior clinical knowledge.