Purpose

Bone age assessment is not only an important means of assessing maturity of adolescents, but also plays an indispensable role in the fields of orthodontics, kinematics, pediatrics, forensic science, etc. Most studies, however, do not take into account the impact of background noise on the results of the assessment. In order to obtain accurate bone age, this paper presents an automatic assessment method, for bone age based on deep convolutional neural networks.

Method

Our method was divided into two phases. In the image segmentation stage, the segmentation network U-Net was used to acquire the mask image which was then compared with the original image to obtain the hand bone portion after removing the background interference. For the classification phase, in order to further improve the evaluation performance, an attention mechanism was added on the basis of Visual Geometry Group Network (VGGNet). Attention mechanisms can help the model invest more resources in important areas of the hand bone.

Result

The assessment model was tested on the RSNA2017 Pediatric Bone Age dataset. The results show that our adjusted model outperforms the VGGNet. The mean absolute error can reach 9.997 months, which outperforms other common methods for bone age assessment.

Conclusion

We explored the establishment of an automated bone age assessment method based on deep learning. This method can efficiently eliminate the influence of background interference on bone age evaluation, improve the accuracy of bone age evaluation, provide important reference value for bone age determination, and can aid in the prevention of adolescent growth and development diseases.

中文翻译：

---

基于深度卷积神经网络与分割相结合的骨龄评估。

目的

骨龄评估不仅是评估青少年成熟度的重要手段，而且在正畸，运动学，儿科学，法医学等领域也起着不可或缺的作用。然而，大多数研究并未考虑背景的影响对评估结果的影响。为了获得准确的骨龄，本文提出了一种基于深度卷积神经网络的骨龄自动评估方法。

方法

我们的方法分为两个阶段。在图像分割阶段，使用分割网络U-Net获取掩模图像，然后将其与原始图像进行比较，以去除背景干扰后获得手骨部分。在分类阶段，为了进一步提高评估性能，在视觉几何组网络（VGGNet）的基础上增加了一种关注机制。注意机制可以帮助模型在手骨重要区域上投入更多资源。

结果

在RSNA2017小儿骨龄数据集上测试了评估模型。结果表明，我们调整后的模型优于VGGNet。平均绝对误差可以达到9.997个月，这优于其他用于骨龄评估的常见方法。

结论

我们探索了基于深度学习的自动骨龄评估方法的建立。该方法可以有效消除背景干扰对骨龄评估的影响，提高骨龄评估的准确性，为骨龄的确定提供重要的参考价值，有助于预防青春期生长发育疾病。