Deep Stacking Ensemble With Explainable AI for Brain Tumor Classification From Multi-Source MRI

Abstract

Gliomas, meningiomas and pituitary adenomas follow different treatment pathways, so subtype classification from MRI carries direct clinical weight. Deep learning methods built on a single backbone and trained on one benchmark generalise poorly once the scanner or the site changes. This paper describes a stacking ensemble that combines EfficientNet-B0 with ResNet-50 through a logistic regression meta-learner. Each base model emits a four-class softmax vector, and the concatenated eight-dimensional descriptor is mapped to a prediction by a meta-learner that adds 36 scalar parameters to a 30.8 M-parameter base. The training corpus merges three public benchmarks, deduplicated by MD5 hash and split 80/10/10 with inverse-frequency class weighting. Four explainable AI components support interpretation: GradCAM++ heatmaps, layer-wise activation ablation, CLAHE edge visualisation and entropy-based out-of-distribution detection. On a held-out test set of 1,275 images the ensemble reaches 95.8% accuracy (95% CI: 94.5–96.7), macro-F1 0.953 and macro-AUC 0.992, improving on both base models in every per-class metric and on five recent published methods in aggregate. We also report deployment complexity, a subtype-level error analysis, and the validation evidence that is still missing.