CatBoost-Stacked Heterogeneous Deep Ensembles for Explainable Multi-Class Brain Tumor MRI Classification

Abstract

Accurate brain tumor screening from MRI requires reliable four-class recognition (glioma, meningioma, pituitary, and no-tumor) under substantial multi-source variability, class imbalance, and the need for clinically interpretable decisions beyond aggregate accuracy. We propose a stacking ensemble that couples heterogeneous feature extractors (EfficientNetB0, MobileNetV2, GoogLeNet, a multi-level CapsuleNet, and a CNN) with a CatBoost meta-learner trained on concatenated class-probability vectors, enabling non-linear fusion of complementary error patterns across backbones. Experiments are conducted on M2, a merged dataset of 24,618 MR images assembled from four public sources, using an 80/5/15 split and stratified 10-fold cross-validation; to prevent evaluation leakage, imbalance mitigation (Borderline-SMOTE and label-preserving augmentation) is applied exclusively to the training folds. The proposed model attains 98.99±0.41% accuracy, 98.32±0.49% micro-F1, 99.33±0.27% PR-AUC, and 94.14±0.13% MCC on M2, consistently outperforming individual backbones, and Grad-CAM audits typically highlight tumor-relevant regions while residual failures concentrate in morphologically similar subtypes. This work contributes a compact, explainable stacking framework with a deployment-oriented inference workflow and visual auditing, validated at scale on heterogeneous MRI sources to support trustworthy brain tumor decision support in resource-constrained settings and real-world clinical practice.