Bone age assessment by multi-granularity and multi-attention feature encoding.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Bone age assessment (BAA) is critical for diagnosing growth disorders and optimizing treatment.
• Current BAA methods suffer from observer variability and low consistency, necessitating automated solutions.

Purpose of the Study:

• To develop an automated bone age assessment method to improve diagnostic accuracy and consistency.
• To introduce the Multi-Granularity and Multi-Attention Net (2M-Net) for enhanced fine-grained feature learning and global understanding.

Main Methods:

• The 2M-Net utilizes a jigsaw method for multi-task learning, generating inductive biases without annotations.
• A hierarchical sharing mechanism trains the model on tasks emphasizing different granularities.
• A self-attention mechanism enhances feature representation, and multi-scale features are used for prediction.

Main Results:

• The 2M-Net was developed and validated on a public dataset of 14,236 hand radiographs.
• The model achieved a mean absolute error (MAE) of 3.98 months compared to reviewer estimates.
• Performance was consistent across sexes, with MAEs of 3.89 months for males and 4.07 months for females.

Conclusions:

• The 2M-Net, incorporating multi-task learning and self-attention, offers a robust automated solution for bone age assessment.
• The developed method demonstrates performance comparable to existing state-of-the-art techniques.