Oral Presentation ANZBMS-MEPSA-ANZORS 2022

Introduction: Knee osteoarthritis is a degenerative joint disease affecting the entire joints including cartilage, ligaments, bone and muscles. We aim to use artificial intelligence (AI, Deep Learning) in the screening of knee osteoarthritis (OA).

Objectives: We developed an automated model for staging knee osteoarthritis severity from radiographs, and we evaluated the prognostic performance of the automated model in the classification of OA.

Methods and Results: This study was a part of the Vietnam Osteoporosis Study (VOS) involved of 1503 women and 934 men aged 40 and older living in Ho Chi Minh city, Vietnam. Radiographs from the VOS staged by radiologists utilised the Kellgren-Lawrence (KL) score. Prior to the usage of those images as the input to a convolutional neural network model, they were standardized and augmented. The model was trained with 1945 plain Xray images, evaluated with 491 images. Applied object detection and other tools of deep learning (Grad-cam maps) were generated to reveal the features utilised (osteophyte formation) by the model to determine KL grades.

Our model was able to achieve 98% in the detection of the exact location of the knee joints and the label of different sides of the knees. The concordance between clinical diagnosis and AI-based classification of knee OA is significant, which the proportion of concordance was able to reach 79.8% (95% CI: 0.77, 0.82). Our model achieved with averagely 91.82% of specificity and 61,52% of sensitivity in all severity levels of knee OA. In terms of screening osteoarthritis, our model achieved 90% (95% CI: 0.87, 0.92) with 94% of sensitivity and 78% of specificity for classifying the knee OA and healthy people.

Conclusions: The AI-based model achieved high sensitivity and specificity in predicting osteoarthritis severity. The model can be used as a screening tool in high-volume setting to reduce clinical workload.