Influence of Site Effects on Radiomics-Based Knee Injury Diagnosis

Impact Magnetic field strength (MFS) substantially affects feature distributions and may introduce bias into machine learning-based disease classification. These findings highlight the need to balance field strength variations in multi-site MRI datasets to enhance model generalization and clinical applicability. Synopsis Motivation: MRNet is a widely used database for knee injury diagnosis. While several classifiers achieve high accuracy on internal datasets, their performance often declines on external data. Researchers primarily attributed this degradation to site effects, though the role of MFS variations has likely been underestimated. Goals: To disentangle the effects of MFS and other site-related factors in knee injury detection. Approach: MFS was inferred using a Gaussian mixture model. Radiomic features were evaluated through one-way ANOVA to quantify site effects. Results: Preliminary results suggest that MFS substantially influences radiomic feature distributions and introduces bias, particularly in imbalanced datasets.