Explainable Deep Learning Framework for Multimodal Brain Tumor Classification via Neuroimaging Attribute Extraction

Accurate grading of brain tumors from multiparametric MRI is a critical step in treatment planning, yet deep learning models trained for this task remain opaque in their reasoning, limiting their acceptance in clinical settings that require transparent accountability. To address this, we propose MADEX, a multimodal attribute-based decision-explanation framework that extracts interpretable neuroimaging attributes using convolutional encoders and organizes them into hierarchical decision trees aligned with the WHO glioma classification structure. An adaptive modality exchange mechanism automatically compensates for artifact-corrupted sequences by drawing on complementary information from higher-quality modalities, and layer-wise Dempster-Shafer evidential fusion then combines modality-specific evidence with uncertainty quantification at each hierarchical level. Prototype constraints further enforce clinical consistency by aligning learned attributes with established diagnostic criteria. Evaluation on the BT-large-2c and BRaTS 2021 datasets shows that MADEX achieves classification accuracies of 89.7% and 87.3%, respectively, while offering superior interpretability through sparse attribute representations, in which just 10 features capture over 90% of the full model's performance. Visualization studies, insertion-deletion tests, and cross-dataset validation confirm that the learned explanations reflect genuine diagnostic features rather than spurious correlations. These properties, together with honest communication of uncertainty and close alignment with clinical reasoning workflows, position MADEX as a viable foundation for responsible AI deployment in neuro-oncological decision support.