X-VIG deep learning framework for automated knee OA grading on plain films

X-VIG, a deep learning model fusing AP and lateral knee radiographs, achieved AUC 0.9646–0.9945 for Kellgren-Lawrence grading across five subregions, outperforming junior radiologists and reducing reading time by 82% (27 s per case) in a reader study.

• Multimodal fusion of AP and lateral views yielded a mean AUC improvement of 5.6 percentage points over single-view models.
• X-VIG matched or surpassed senior radiologists in most subregions while cutting reading time by 82% (27 seconds per case).
• Grad-CAM++ visualizations provide interpretable heatmaps to support clinical acceptance.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.