Transfer Learning From Hand-Trained Deep Learning Models to Estimate Bone Age From Knee Radiographs

Area of Science:

• Orthopaedic surgery
• Radiology
• Artificial Intelligence

Background:

• Accurate skeletal age assessment is crucial for pediatric orthopaedic care.
• Traditional methods like the Greulich and Pyle atlas require additional hand imaging and radiation.
• Knee imaging is often readily available, presenting an opportunity for alternative assessment methods.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for estimating bone age directly from knee radiographs.
• To provide a more efficient and less radiation-intensive method for skeletal maturity assessment in skeletally immature patients.

Main Methods:

• A ConvNeXT deep learning model was trained on 2374 knee radiographs from patients aged 18 years or younger.
• The dataset included paired hand films for ground-truth bone age determination.
• Model performance was evaluated using Mean Absolute Error (MAE) and Bland-Altman analysis, with gradient-based saliency maps for interpretability.

Main Results:

• The DL model achieved a Mean Absolute Error (MAE) of 5.02 months, significantly outperforming the abbreviated Fels method (9.59 months).
• An even lower MAE of 3.43 months was achieved using pseudo-labels from a hand DL model.
• Bland-Altman analysis indicated excellent agreement between the model's predictions and the ground-truth bone ages.

Conclusions:

• Automated bone age estimation from knee radiographs using deep learning is feasible and highly accurate.
• This AI-powered tool can assist orthopaedic surgeons and radiologists in clinical decision-making for skeletally immature patients.
• Further external validation and refinement are recommended for widespread clinical adoption.