CMDA-Net: Auditable SHAP-based brain-tumor MRI classification, with a cross-modal dynamic-attention design for modality-robust segmentation

Reliable computer-aided analysis of brain-tumor MRI must be accurate, robust to the imaging conditions encountered in practice, and interpretable in a form that can be documented and reviewed. We present CMDA-Net, a framework built around two contributions. The first, which we validate experimentally, is an auditable classification pipeline: a 2D residual–CBAM classifier (CMDA-Cls) coupled with a SHAP-based interpretation layer that goes beyond heatmaps to emit a centre-versus-periphery separability metric and an attribution-weighted confusion matrix, archived in a versionable multi-sheet report to support post-hoc auditing. On the four-class Brain Tumor MRI benchmark, CMDA-Cls attains $$99.47\%$$ accuracy and $$99.44\%$$ macro-F1 on the full held-out test set ( $$N{=}1311$$ ), with $$97.09\%\!\pm \!0.55\%$$ accuracy under stratified 5-fold cross-validation, and the centre-to-periphery attribution ratio exceeds one for every class (from 1.17 on pituitary to 2.94 on meningioma), excluding peripheral-shortcut behaviour. The second contribution is architectural: a 3D multi-modal segmentation network (CMDA-Seg) that couples Cross-Modal Attention Fusion (CMAF), a survival-constrained Modality-Dropout (MoDrop) regularizer and a FiLM-conditioned decoder to fuse MRI modalities adaptively and to degrade gracefully when modalities are missing. Because a multi-parametric segmentation cohort was not available for this study, CMDA-Seg is presented as a design contribution whose four-modality evaluation is left to future work.