Generative AI for KL Grading: A Few-Shot Image Augmentation Approach

Deep learning in orthopedic medical image analysis has gained momentum in recent years, but progress has been hindered by the lack of large-scale, standardized ground-truth imaging datasets. This study introduces an innovative deep few-shot image augmentation pipeline designed to synthetically generate high-fidelity knee radiographs, specifically aimed at improving downstream tasks such as Kellgren-Lawrence (KL) grading for knee osteoarthritis. Despite limited training data, the proposed method successfully generates realistic plain knee radiographs, which are then used to train a KL grade classifier. Experimental results demonstrate strong performance: the synthetic images achieved average Frechet Inception Distance (FID) scores of 26.33 for KL grading and 22.538 for bilateral knee views. The classifier trained on these images reached a test accuracy of 72.7% and a Cohen’s Kappa score of 0.451. In addition to its technical success, the study contributes a publicly available dataset of 86,000 synthetic knee radiographs, offering valuable resources to the research community. Overall, the proposed pipeline effectively addresses data scarcity in orthopedic AI, paving the way for broader applications in medical image analysis and AI-driven diagnostic support.

GitHub: GAN_Synthetic_Knee_Radiographs

Paper: Published in JAMIA