Determination of Modified Waldenström Staging in Legg-Calvé-Perthes Disease Using Deep Learning

Area of Science:

• Orthopaedic imaging analysis
• Artificial intelligence in medicine
• Pediatric skeletal conditions

Background:

• Legg-Calvé-Perthes disease (LCPD) management relies on age and disease stage.
• The modified Waldenström system is standard but has moderate inter-rater reliability.
• Deep learning (DL) offers potential for consistent, efficient image-based classification in orthopaedics.

Purpose of the Study:

• To develop a deep learning (DL) model for determining modified Waldenström staging of LCPD.
• To assess the DL model's accuracy in classifying LCPD stages from standard hip radiographs.

Main Methods:

• A DL classification pipeline was developed using AP and frog-lateral hip radiographs from LCPD patients.
• The model extracted quantifiable epiphyseal parameters.
• Performance was evaluated on early (Ia-IIa) vs. late (IIb-IV) LCPD classification and the full staging system.

Main Results:

• The DL pipeline achieved excellent segmentation (dice coefficient = 0.93) and rapid parameter extraction (12s).
• On hold-out testing, the model showed an AUROC of 0.82 for early vs. late LCPD classification.
• External validation yielded an AUROC of 0.75 for early vs. late stage classification.

Conclusions:

• The DL model accurately differentiates early vs. late LCPD stages from radiographs, with moderate accuracy for full staging.
• Automated segmentation provides objective assessment of key femoral morphology parameters.
• AI models can standardize Waldenström staging in clinical practice and research, though larger, diverse datasets are needed for improved generalizability.