Abstract: Bone age assessment is a common clinical method to study endocrine,genetic factors and growth disorders in children.The traditional bone age assessment methods are time-consuming,and are vulnerable to the effects of subjective factors of the evaluator.The existing bone age assessment methods that automatically extract clinical features have low accuracy and poor generalization ability.This paper proposes an end-to-end automatic bone age assessment method based on deep learning for X-ray images of hand bones.The method removes the Softmax layer to optimize the structure of Inception ResNet V2 network,and adds asymmetric convolution kernel in the Inception module to improve the classification accuracy of feature graph.In addition,it introduces residual connection structure to avoid gradient disappearance or explosion problems,and the Mean Square Error(MSE) loss function is used to evaluate the regression performance of bone age assessment.A stratified K-fold cross-validation method is used to ensure the balanced classification of samples in dataset.Experimental results show that compared with BoNet-based bone age assessment methods,the Mean Absolute Error(MAE) between the estimated bone age and the true bone age was reduced by 0.423 0 years,and its accuracy of bone age prediction is higher.

Key words: bone age assessment, deep learning, X-ray image, stratified K-fold cross-validation method, Inception ResNet V2 network

摘要： 骨龄评估是研究儿童内分泌、遗传因子和生长障碍的常用临床手段，传统骨龄评估方法耗时较长，易受评估者主观因素影响产生误差，而现有自动提取临床特征的骨龄评估方法精度低且泛化能力差。提出一种基于深度学习的端到端手骨X射线图像自动骨龄评估方法。去除Inception ResNet V2网络的Softmax层优化结构，在Inception模块中增加非对称卷积核提高特征图分类精度，引入残差连接结构避免梯度消失或爆炸问题，同时采用均方误差损失函数对骨龄评估回归性能进行评价，并使用分层K折交叉验证法保证数据集样本分类均衡。实验结果表明，与采用BoNet网络的骨龄评估方法相比，该方法评估的骨龄与真实骨龄平均绝对误差减少0.423 0岁，骨龄预测精度更高。

关键词: 骨龄评估, 深度学习, X射线图像, 分层K折交叉验证法, Inception ResNet V2网络

CLC Number: 

• TP391