Study on multimodal spatially-constrained contrastive learning for knee osteoarthritis severity grading.

Area of Science:

• Biomedical Engineering
• Orthopedics
• Machine Learning

Background:

• Knee osteoarthritis (KOA) classification faces challenges with single-modal data and imbalanced datasets.
• Accurate KOA assessment is crucial for effective treatment and management.

Purpose of the Study:

• To develop a novel Multimodal Spatial-constraint Contrastive Learning (MSCL) model for precise KOA severity stratification.
• To overcome limitations in feature coverage and class distribution imbalance in KOA classification.

Main Methods:

• Synchronous acquisition of dynamic/static plantar pressure and human keypoint trajectory data.
• Utilizing graph convolutional networks, Transformers, and Cross Attention for multimodal spatial-temporal fusion.
• Employing a pyramid CNN for spatial constraint generation and contrastive learning for KL grading.

Main Results:

• The MSCL model achieved a 0.94 macro-average accuracy in Kellgren-Lawrence (KL) grading.
• Demonstrated a 7% improvement in F1-scores for imbalanced KOA categories with limited samples.
• Established a novel paradigm for KOA assessment via multimodal gait analysis.

Conclusions:

• The MSCL model offers a robust and accurate approach for KOA severity stratification.
• Multimodal gait analysis combined with contrastive learning enhances classification performance.
• This study provides a significant advancement in the objective assessment of knee osteoarthritis.