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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Endocrinology

Background:

• Bone age assessment (BAA) is crucial for evaluating skeletal maturity and identifying growth discrepancies.
• Manual BAA methods are complex and require expert interpretation.
• Existing automated BAA techniques struggle with small, imbalanced datasets, a common challenge in deep learning.

Purpose of the Study:

• To develop an improved deep learning-based method for bone age assessment.
• To address the limitations of small and imbalanced sample sizes in automated BAA.
• To enhance the accuracy and efficiency of skeletal age determination.

Main Methods:

• Proposed a novel deep learning framework incorporating pre-training and advanced training architectures.
• Introduced a data augmentation technique using cosine distance within an optimal transport framework (CNN-GAN-OTD).
• Investigated different training strategies, including label order (gender and bone age) and supervised/semi-supervised learning.

Main Results:

• The optimal training architecture combined CNN-GAN-OTD data augmentation with a supervised Inception v3 network using a gender-last classification approach.
• This configuration achieved the best bone age assessment performance with a mean average error of 4.23 months.
• The study demonstrated the effectiveness of the proposed data augmentation and training strategy.

Conclusions:

• The developed CNN-GAN-OTD data augmentation framework shows potential as a versatile component for general deep learning networks.
• The findings on label order in multi-label tasks offer insights for future deep learning model development.
• The study provides a more accurate and robust automated solution for bone age assessment.