Knee Osteoarthritis Severity Grading Using Contrastive Learning Image Pre-Training

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging

Background:

• Accurate knee osteoarthritis (KOA) severity grading is crucial for patient care.
• Manual radiographic grading methods are prone to observer variability.
• Developing objective tools for KOA assessment is essential.

Purpose of the Study:

• To evaluate a fine-tuned contrastive language-image pre-training (CLIP) framework for grading KOA severity.
• To assess the model's performance using the Kellgren-Lawrence (KL) classification system (Grades 0-4).
• To determine the utility of AI in assisting clinicians with KOA radiographic evaluation.

Main Methods:

• Utilized a CLIP framework projecting visual features from radiographs and textual descriptions into a shared embedding space.
• Trained the model on 8260 posterior-anterior (PA) fixed-flexion X-ray images from the Osteoarthritis Initiative (OAI) dataset.
• Externally validated the model on an independent dataset of 1650 plain radiographs.

Main Results:

• The fine-tuned CLIP model achieved 76.94% accuracy and a 76.66% F1-score on the external validation dataset.
• Demonstrated competitive and stable diagnostic capabilities across distinct data distributions.
• Highlighted the effectiveness of aligned vision-language representations for KOA grading.

Conclusions:

• Fine-tuned CLIP architectures provide a robust foundation for computer-aided KOA evaluation.
• The developed framework offers semantically transparent diagnostic assistance.
• This AI approach has the potential to enhance the objectivity and consistency of KOA severity assessment.