Fig. 3

The flowchart of the radiomics model construction. A total of 951 features were retained based on ICC ≥ 0.75. To address class imbalance, upsampling was applied to the training set, followed by min–max normalization for standardization. Feature reduction using the Pearson correlation coefficient (threshold 0.99) reduced the feature set to 656, which was further refined to 2 features using Analysis of Variance (ANOVA). Gradient boosting classifiers were then trained on these optimized features using the balanced training set