Knee osteoarthritis (KOA) is a progressive joint disorder that requires early and accurate diagnosis for effective treatment. This study explores the use of deep learning to classify KOA severity from knee X-ray images. Several convolutional neural network (CNN) architectures were trained and evaluated to assess their performance in distinguishing different severity levels. While some models demonstrated strong classification accuracy, others faced challenges in identifying early-stage KOA. The results highlight the potential of AI-driven approaches in medical diagnostics, emphasizing the need for further model refinement, improved data preprocessing, and advanced regularization techniques to enhance predictive performance and clinical applicability. The proposed study involved models: CoAtNet, ConvNeXt, Inception-ResNet, MultiResUNet, and Swin-UNet Encoder. The results show that the custom CNN model achieved the highest classification accuracy of 88%, outperforming other models in precision and recall for most classes.