Spatially Normalized X-ray Imaging Using Landmark-based Registration for Knee Osteoarthritis Severity Classification

Knee osteoarthritis (OA) is commonly assessed from X-ray images using the Kellgren-Lawrence (KL) grading system. However, variations in imaging protocols and patient positioning introduce spatial inconsistencies that may affect the performance of convolutional neural network (CNN)-based models. This study proposes a landmark-based spatial normalization framework to align anatomical structures before classification, thereby reducing geometric variability in an ordinal learning setting. The method is evaluated across multiple CNN architectures on the Osteoarthritis Initiative (OAI) dataset and further assessed through cross-dataset testing on the Kaggle Osteoarthritis dataset. The results show consistent, moderate improvements, with the best-performing ConvNeXt-Tiny model increasing QWK from 0.852 to 0.857 on the OAI dataset and from 0.941 to 0.943 on the Kaggle Osteoarthritis dataset. Robustness analysis under controlled spatial perturbations indicates relatively consistent performance, with QWK remaining above 0.80 at moderate levels of noise. These findings suggest modest improvements in feature alignment and prediction consistency within the evaluated settings.