ReSHAPe: A redundancy-reduced SHAP-based feature selection pipeline for interpretable radiomics in biomedical image analysis

Handcrafted radiomic descriptors are widely used in biomedical image analysis, but classic radiomics pipelines often suffer from high feature redundancy and an underdeveloped, weakly principled feature-selection practice, which together can impair generalization and limit model interpretability. To address this, we introduce ReSHAPe (Redundancy-Reduced SHAP-based Evaluation), a two-stage, model-aware feature selection pipeline that makes SHAP-driven selection practical for radiomics. ReSHAPe first performs redundancy pruning by removing highly correlated features using Spearman rank correlation, retaining within each correlated group the descriptor with the lower absolute skewness. It then applies SHAP-based global importance to rank the remaining features and iteratively select a compact subset; an ensemble variant aggregates SHAP rankings across multiple classifiers to promote consensus and interoperability. We evaluate ReSHAPe on three MedMNIST v2 subsets (BreastMNIST, PneumoniaMNIST, BloodMNIST) using 285 handcrafted features and five well-known classifiers (SVM, Decision Tree, Random Forest, Extra Trees, XGBoost), comparing against univariate filters (ANOVA F-test, mutual information), SHAP-only selection, correlation-based filtering, and full-feature baselines. Across datasets, ReSHAPe preserves performance while drastically reducing dimensionality; on radiomic tasks, it is consistently competitive with SHAP-only selection f-measure weighted values differences typically lower than 0.03, and it remains effective in the non-radiomic multiclass setting (maximum decrease of f-measure weighted value lower than 0.04). Finally, the correlation pre-filtering stage markedly reduces SHAP overhead, which would otherwise require 200 additional model training/evaluation steps when applied directly to the full feature space.