Generative adversarial network based data augmentation and gender-last training strategy with application to bone age assessment

Bone age assessment (BAA) is widely used in determination of discrepancy between skeletal age and chronological age. Manual approaches are complicated which require experienced experts, while existing automatic approaches are perplexed with small and imbalanced samples which is a big challenge in deep learning.

In this study, we proposed a new deep learning based method to improve the BAA training in both pre-training and training architecture. In pre-training, we proposed a framework using a new distance metric of cosine distance in the framework of optimal transport for data augmentation (CNN-GAN-OTD). In the training architecture, we explored the order of gender label and bone age information, supervised and semi-supervised training.

We found that the training architecture with the CNN-GAN-OTD based data augmentation and supervised gender-last classification with supervised Inception v3 network yielded the best assessment (mean average error of 4.23 months).

The proposed data augmentation framework could be a potential built-in component of general deep learning networks and the training strategy with different label order could inspire more and deeper consideration of label priority in multi-label tasks.