Self-Supervised Attention Mechanism for Pediatric Bone Age Assessment With Efficient Weak Annotation

Area of Science:

• Medical Imaging and Artificial Intelligence
• Pediatric Endocrinology
• Radiology

Background:

• Pediatric bone age assessment (BAA) is crucial for diagnosing growth and endocrine disorders.
• Current deep learning methods for BAA rely on precise annotations of anatomical Regions of Interest (RoIs), limiting clinical applicability.
• The need for automated RoI discovery in BAA is significant for enhancing clinical utility.

Purpose of the Study:

• To develop a novel self-supervised learning mechanism for automatic discovery of informative RoIs in BAA.
• To overcome the limitations of precise RoI annotations in existing deep learning-based BAA methods.
• To propose an end-to-end BAA method utilizing only image-level annotations.

Main Methods:

• Introduced PEAR-Net (Part Extracting and Age Recognition Network), a self-supervised model for BAA.
• Employed a Part Extracting (PE) agent for automatic RoI discovery and an Age Recognition (AR) agent for age assessment.
• Utilized self-consistency of RoIs to optimize the PE agent, enabling mutual reinforcement between PE and AR agents.

Main Results:

• Achieved state-of-the-art performance on the RSNA 2017 dataset for pediatric bone age assessment.
• Reported a Mean Absolute Error (MAE) of 3.99 months, demonstrating high accuracy.
• Successfully demonstrated automatic RoI discovery using only image-level annotations.

Conclusions:

• PEAR-Net effectively discovers informative RoIs for BAA without requiring precise annotations.
• The proposed self-supervised approach enhances the clinical value of deep learning in pediatric BAA.
• This work represents the first end-to-end BAA method capable of automatic RoI discovery with image-level supervision.