Automatic Detection of Knee Osteoarthritis Severity with SOTA Deep Learning Models and Ordinal Loss

Knee Osteoarthritis (OA) is one of the most common bone diseases that causes pain, stiffness, and limited mobility due to the damage to cartilage. The diagnosis of the disease is made by medical doctors based on the individual's symptoms and findings on X-ray. During this process, different evaluations among medical doctors usually based on experience can make the diagnosis of the disease difficult. In this paper, deep learning models were used for automatic detection of knee OA stages according to the Kellgren-Lawrence (KL) grading system in order to facilitate the disease diagnosis process. Pre-trained SOTA deep learning models including ConvNeXt and ConvNeXt V2 are fine tuned using the ordinal loss function. This loss function imposes a penalty, taking into account the distance difference between the real class and the predicted class, in cases where there is a order between the labels, like KL grades that representing OA severity. In this study that we classified the knee OA stage according to KL grades, we obtained the best performances with 73.91% accuracy, 73.77% F1 score and 0.28 mean absolute error (MAE) values on test set obtained on ConvNeXt V2-Tiny and ConvNeXt-Base models that we trained with the ordinal loss function.