Explainable multimodal knee osteoarthritis diagnosis from X-ray images using anomaly detection with clinical and structural biomarkers.

Knee Osteoarthritis (KOA) is a common degenerative joint disease that impairs mobility and affects patients' quality of life. Although many studies have applied AI in diagnosing KOA with some successes, numerous challenges remain unresolved. Most current diagnostic models primarily rely on X-ray images without incorporating clinical factors and knowledge, thereby limiting accuracy, especially at early stages. Moreover, existing models often operate as "black boxes," providing results without explaining the rationale, restricting their practical application due to transparency requirements in healthcare. In response, we introduce the xDesCO framework, a novel explainable AI model that not only delivers accurate diagnostic results by integrating clinical knowledge but also offers visual explanations for diagnoses. As a predictive model, xDesCO integrates three clinical factors-disease indicators, Joint Space Width, and clinical data-to provide more accurate assessments, achieving 73.79% accuracy on the public OAI dataset, outperforming other methods. Using a novel ROI-masking generative adversarial model, xDesCO offers an anomaly map that highlights abnormalities in a human-readable visual format, enhancing transparency and usability. Furthermore, we developed a web app enabling users to upload X-rays and promptly receive diagnostic results and explanations. Our framework sets a new standard for explainable and effective KOA diagnosis.