Modality-level attribution and redundancy-aware radiomics for MRI-based differentiation of melanoma and NSCLC brain metastases.

Background: Differentiating the primary source of brain metastases using imaging alone is difficult, especially when intracranial disease is the initial sign of cancer. This study aimed to develop and validate a redundancy-aware, explainable radiomics framework for distinguishing melanoma from non-small cell lung cancer (NSCLC) brain metastases through multiparametric MRI. Methods: Lesion-level radiomic features were extracted from T1, contrast-enhanced T1 (T1CE), T2, and FLAIR sequences. Correlation-based redundancy filtering and L1-regularized feature ranking were applied to minimize collinearity among features. A structured parametric analysis across progressively larger top-K feature subsets was performed to evaluate performance stability and modality-specific contribution patterns. Multiple machine learning classifiers were tested using cross-validation and an independent test set, with model interpretability assessed via TreeSHAP to quantify feature- and modality-level contributions. Results: The best-performing configuration achieved a test AUC of 0.75, while RF-200 was selected for the primary tree-based explainability analysis due to its stable AUC-based discrimination. SHAP analysis showed that classification depended on a small set of predictors combining texture and intensity distribution features. FLAIR and T2 features contributed most to model attribution, suggesting that fluid-sensitive sequences capture discrimination-relevant heterogeneity and microenvironmental signals beyond contrast enhancement. Discussion: A redundancy-aware, interpretable radiomics approach shows promising preliminary discrimination between melanoma and NSCLC brain metastases and provides structured insight into modality-specific sequence contributions. External validation and prospective reader studies are required before clinical translation.