Can a Convolutional Neural Network Classify Knee Osteoarthritis on Plain Radiographs as Accurately as Fellowship-Trained Knee Arthroplasty Surgeons?

Abstract

Osteoarthritis (OA) is the leading cause of disability among adults in the United States. As the diagnosis is based on the accurate interpretation of knee radiographs, use of a convolutional neural network (CNN) to grade OA severity has the potential to significantly reduce variability. Knee radiographs from consecutive patients presenting to a large academic arthroplasty practice were obtained retrospectively. These images were rated by 4 fellowship-trained knee arthroplasty surgeons using the International Knee Documentation Committee (IKDC) scoring system. The intraclass correlation coefficient (ICC) for surgeons alone and surgeons with a CNN that was trained using 4755 separate images were compared. Two hundred eighty-eight posteroanterior flexion knee radiographs (576 knees) were reviewed; 131 knees were removed due to poor quality or prior TKA. Each remaining knee was rated by 4 blinded surgeons for a total of 1780 human knee ratings. The ICC among the 4 surgeons for all possible IKDC grades was 0.703 (95% confidence interval [CI] 0.667-0.737). The ICC for the 4 surgeons and the trained CNN was 0.685 (95% CI 0.65-0.719). For IKDC D vs any other rating, the ICC of the 4 surgeons was 0.713 (95% CI 0.678-0.746), and the ICC of 4 surgeons and CNN was 0.697 (95% CI 0.663-0.73). A CNN can identify and classify knee OA as accurately as a fellowship-trained arthroplasty surgeon. This technology has the potential to reduce variability in the diagnosis and treatment of knee OA.