Detection of knee osteoarthritis severity from x-ray images using transfer learning methods

Knee osteoarthritis (KOA) is a chronic and progressive joint disease characterized by the deterioration of cartilage tissue in the joints. Appropriate treatment and early diagnosis are important for disease control. However, traditional diagnostic methods used to classify KOA from X-ray images require expertise and, unfortunately, have a large margin of error. This study presents an image processing-based solution for detecting KOA severity from X-ray images using the Bilateral filter, contrast-limited adaptive histogram equalization (CLAHE), and transfer learning models. While the CLAHE method improved the image quality, the Bilateral filter improved the details in X-ray images and minimized blurring. The KOA image dataset consists of 9786 knee images and five class labels. In addition, in the scope of this study, the performances of AlexNet, VGG19, DenseNet201, EfficientNetB0, and ResNet101 transfer learning and Bilateral, Wiener, Gauss, Sobel, and Laplacian image filtering techniques were compared. ResNet101 transfer learning model with Bilateral image filtering achieved the highest success with kappa statistics of 0.970, a weighted F1 score of 0.978, and an accuracy of 97.85%. In conclusion, this study highlights the importance of filter selection in medical image classification tasks and the diagnostic potential of combining it with transfer learning models.