Knee osteoarthritis (KOA) is a prelevent disease associated with serious adverb effects, and its diagnosis requires significant time and effort of radiologists. We proposed applying deep learning for the KOA problem. To aid in the diagnosis of KOA, this study conducted experiments on six deep learning models, including ResNet50, ResNet101, ResNet152, SEResNet50, SEResNet101, and SEResNet152, using the transfer learning technique, with the KOA Severity Grading Dataset, which contained X-ray images classified according to the Kellgren and Lawrence (KL) system. The study focused on two classification tasks, namely binary classification, in which images were labeled Negative or Positive to detect KOA, and severity classification, in which images were classified based on the severity levels of KOA. To facilitate this analysis, the original dataset was structured into three different configurations. The first dataset retained five classes of KL grades, and the second dataset merged the Healthy and Doubtful classes into Negative class, while the Minimal, Moderate and Severe classes were merged into Positive class, and the third dataset had three classes, including Healthy (combining Healthy, Doubtful, Minimal), Moderate, and Severe. The highest classification accuracies acheved for these datasets were 0.6926 (SEResNet50), 0.8339 (ResNet50), and 0.9493 (SEResNet152), respectively. The results indicate that deep learning models are well suited for analyzing KOA X-ray images. Depending on the classification task, different models achieve the highest accuracy, highlighting the importance of selecting the most appropriate model for each specific problem.