基于深度学习实现肋软骨CT图像自动化骨龄推断

Automatic bone age estimation of costal cartilage CT Images based on deep learning

目的 利用深度学习方法对第1~7对成人肋软骨CT重建图像进行特征提取,实现成人肋软骨骨龄的自动化推断。方法 回顾性收集年龄在20岁~70岁之间的男性和女性样本各625例,并通过容积再现技术(VRT)得到对应的VRT图像。通过图像预处理和数据增强之后,将其中的500例作为训练集,125例作为测试集,运用5折交叉验证的方法分别评估Res Net、Res Ne Xt、Dense Net及Google Net网络模型的性能,并将5折交叉验证结果的平均值作为最终推断结果。结果 Res Net50网络模型在男性和女性数据集中都取得了最佳实验结果,平均绝对误差分别为4.56岁和3.91岁,±5.0岁范围内预测准确率分别达到64.00%和70.88%,±10.0岁范围内预测准确率分别达到88.96%和94.40%。结论 与传统方法和机器学习方法相比,深度学习方法能够避免人为因素的影响,并且大大提高了成人肋软骨骨龄推断的准确率、降低了预测年龄与真实年龄之间的平均绝对误差,具有较高的临床应用价值。

Objective To use the deep learning methods to extract features of the 1st to 7th adult costal cartilage CT reconstruction images to realize the automatic estimation of adult costal cartilage bone age.Methods 625 male and 625 female samples aged between 20 and 70 years old were collected retrospectively,and the corresponding VRT images were reconstructed by volume rendering technology(VRT).After image preprocessing and data augmentation,500 cases were used as the training set and 125 cases as the test set.The performance of ResNet,ResNeXt,DenseNet and GoogleNet networks was evaluated by using 5-fold cross-validation,and the average value of 5-fold cross-validation results was taken as the final estimation result.Results The ResNet50 network achieved the best results in both male and female datasets.The mean absolute error was 4.56 years and 3.91 years,the accuracy rate was 64.00%and 70.88%in the range of±5.0 years,88.96%and 94.40%in the range of±10.0 years,respectively.Conclusion Compared with traditional methods and machine learning methods,the deep learning models can avoid the influence of human factors,greatly improve the accuracy of adult costal cartilage bone age estimation,and reduce the error between predicted age and real age,which has high clinical application value.