Fully Automated Deep Learning System for Bone Age Assessment.

Area of Science:

• Pediatric endocrinology
• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Bone age assessments (BAA) are crucial for evaluating pediatric endocrine and metabolic disorders.
• Current BAA methods are manual, time-consuming, and have seen little innovation since 1950.
• Accurate BAA is essential for diagnosing and managing various pediatric conditions.

Purpose of the Study:

• To develop and validate a fully automated deep learning pipeline for performing bone age assessments.
• To enhance the accuracy and efficiency of BAA in pediatric patients.
• To provide a decision support tool for clinicians, reducing interpretation time.

Main Methods:

• A deep learning pipeline was created, including automated region of interest segmentation, radiograph standardization, and preprocessing.
• An ImageNet pre-trained, fine-tuned convolutional neural network (CNN) was employed for BAA.
• The input occlusion method was used to generate attention maps for model interpretability.

Main Results:

• The automated BAA system achieved accuracies of 57.32% (female) and 61.40% (male) on held-out test images.
• High accuracy was observed within 1-2 years: 90.39-98.11% (female) and 94.18-99.00% (male).
• Attention maps confirmed the model focuses on features relevant to human expert BAA.

Conclusions:

• The fully automated deep learning system offers a more accurate and efficient approach to BAA.
• The system was successfully deployed in a clinical setting as a decision support tool.
• This AI-driven BAA significantly reduces interpretation time (<2 seconds) compared to manual methods.