Abstract—Accurate and interpretable classification of brain tumors from magnetic resonance imaging (MRI) is a critical step toward computer-aided diagnosis, yet single-model convolutional neural networks (CNNs) often trade off accuracy, robustness, and interpretability. This paper presents an ensemble deep learning framework for four-class brain tumor MRI classification (glioma, meningioma, pituitary tumor, and no tumor) that combines three ImageNet-pretrained backbones — MobileNetV2, EfficientNetB3, and ResNet50V2 — fine-tuned independently via transfer learning and fused through a learned meta-learner (stacked generalization) rather than fixed-weight averaging. Class imbalance handling is built into the pipeline via inverse-frequency class weighting with an additional manual boost for the meningioma class, and inference is stabilized with eight-pass test-time augmentation (TTA). Interpretability is provided through per-backbone Grad-CAM heatmaps fused into a confidence-weighted ensemble heatmap In the experimental evaluation conducted on a 5,600-image, four-class balanced MRI dataset, the individual backbones reached 96.02% (MobileNetV2), 96.38% (EfficientNetB3), and 98.55% (ResNet50V2) validation accuracy, and the stacked meta-learner ensemble achieved 99% accuracy with a weighted F1-score of 0.99. The proposed framework was trained using separate training and testing subsets and evaluated on an independent held-out test set, demonstrating strong generalization capability. The complete pipeline is exposed through an interactive Gradio web application returning a diagnosis, a confidence tier, per-model vote breakdown, and visual explanations. Index Terms—Brain Tumor Classification, MRI, Convolutional Neural Networks, Transfer Learning, Ensemble Learning, Stacked Generalization, Grad-CAM, Test-Time Augmentation, Explainable AI
N et al. (Fri,) studied this question.