Large language models for efficient whole-organ MRI score-based reports and categorization in knee osteoarthritis

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Osteoarthritis Research

Background:

• Knee osteoarthritis (OA) diagnosis and management rely on detailed MRI analysis.
• Manual generation of structured Whole-Organ MRI Score (WORMS)-based reports is time-consuming.
• Accurate OA severity prediction from MRI is crucial for treatment planning.

Purpose of the Study:

• To assess the performance of large language models (LLMs) in generating WORMS-based knee MRI reports.
• To evaluate LLMs' ability to predict knee OA severity from MRI data.
• To compare LLM-generated reports with original reports in terms of surgeon preference and efficiency.

Main Methods:

• 160 knee MRI reports from patients suspected of OA were analyzed.
• GPT-4o and GPT-4o-mini were used with in-context knowledge (ICK) and chain-of-thought (COT) prompting.
• Orthopedic surgeons reviewed and compared LLM-generated reports against original reports.

Main Results:

• GPT-4o achieved 100% accuracy in extracting knee laterality.
• GPT-4o significantly outperformed GPT-4o-mini in WORMS report generation (93.9% vs 76.2%) and OA severity prediction (98.1% vs 68.7%).
• Surgeons preferred LLM-generated reports, found them easier to use, and spent less time reviewing them.

Conclusions:

• GPT-4o can generate expert, multi-feature WORMS-based knee MRI reports from free-text inputs.
• LLM-generated reports enhance clinical workflow efficiency and reduce documentation burden.
• Integration of LLMs shows promise for improving productivity in knee OA management.