Model Methodology and Limits

The system estimates bone age and gender from the hand-wrist radiograph. Gender is not requested from the user; it is an output inferred from the image.

Selected Specialist Workflow

Before analysis, the user selects whether the patient is under 18 or 18 and older. The system then runs the matching specialist model.

  1. Minor specialist: for under-18, estimates bone age in months using a divide-and-conquer ensemble.
  2. Adult specialist: for 18 and older, estimates age in years.

Architecture and Techniques

The models are multi-task convolutional networks. Age is learned as a probability distribution, attention pooling focuses on age-informative hand regions, and horizontal flip test-time augmentation improves stability.

Validity Scope

Expected performance applies to well-positioned hand-wrist radiographs similar to the training and validation data. Atypical anatomy, implants, artifacts, image text overlays, or incorrect positioning may affect the result.

Uncertainty

The minor model is expected to vary at month scale, while adult estimates can vary at year scale because hand bones provide less distinguishing signal after maturity around 18-19 years.