Explainable, Uncertainty-Aware Ensemble Learning for Brain Tumor Classification from MRI with a Simulated Fairness Audit

Most brain MRI classifiers stop at accuracy; whether their explanations hold up, how sure each call is and whether some patients fare worse is rarely checked. We combined five convolutional networks (Xception, InceptionV3, EfficientNetB4, DenseNet121 and ConvNeXtTiny) and a ViT-B/16 transformer through a logistic regression meta-learner fitted on Out-of-Fold predictions. The models were trained with five-fold cross-validation on 5,040 of 7,200 class-balanced MRI images and tested once on 1,080 held-out images. The MetaEnsemble reached 98.33% accuracy, an AUC-ROC of 0.9989 and an Expected Calibration Error of 0.0083, whereas every single model was underconfident. We set LIME, SHAP, Grad-CAM and Grad-CAM + + side by side and scored the Grad-CAM maps (localization accuracy, Pointing Game accuracy, AOPC faithfulness) instead of judging them by eye. Forty Monte Carlo Dropout passes per image gave each case an uncertainty estimate. Lacking patient metadata, we simulated age and sex labels from epidemiological priors to demonstrate a fairness audit; the gaps were small but pointed the same way in every model (age 1.55% to 4.00%, sex 1.07% to 2.05%). The evidence comes from one public dataset, but it shows that explanation scoring, uncertainty estimation and subgroup checks can be planned into a brain tumor classifier from the start.