Multi-Branch Attention Learning for Bone Age Assessment with Ambiguous Label.

Area of Science:

• Pediatric Endocrinology
• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Bone age assessment (BAA) is crucial for diagnosing pediatric endocrine and metabolic disorders.
• Current deep learning BAA models, trained on Western data (RSNA), exhibit limited accuracy for East Asian populations due to developmental and standardization differences.
• Acquiring large, accurately labeled pediatric bone age datasets is challenging and labor-intensive.

Purpose of the Study:

• To develop an accurate deep learning model for bone age assessment in East Asian children.
• To address the limitations of existing models trained on Western datasets.
• To leverage ambiguous radiological labels for improved model training efficiency.

Main Methods:

• Collected a novel bone age dataset (CNBA) from East Asian populations.
• Utilized ambiguous labels from radiology reports, transforming them into Gaussian distribution labels.
• Proposed the multi-branch attention learning with ambiguous labels network (MAAL-Net), incorporating hand object localization and attention-based feature extraction modules.
• Identified informative regions of interest (ROIs) using image-level labels.

Main Results:

• MAAL-Net achieved competitive performance against state-of-the-art methods on both RSNA and CNBA datasets.
• The model demonstrated performance on par with experienced physicians in pediatric bone age assessment.
• The proposed method effectively handles ambiguous labels for robust BAA.

Conclusions:

• MAAL-Net offers a promising solution for accurate and efficient bone age assessment in East Asian children.
• The approach of using ambiguous labels and multi-branch attention learning can overcome data scarcity and improve model generalizability.
• This work contributes to advancing AI-driven diagnostic tools in pediatric endocrinology.