A Meta-Learner Driven CNN–Transformer Stacked Ensemble for High-Accuracy Brain Tumor Classification

Abstract

Diagnosis of brain tumor with Magnetic Resonance Imaging (MRI) is the key for early treatment planning, which however, manual interpretation of can be time-consuming and may suffer from diagnostic inconsistency. To overcome this issue, in this paper, a stacked ensemble-based hybrid deep learning framework for multi-class brain tumor classification is proposed. The combination consists of Xception, ConvNeXt and Swin Transformer as heterogeneous base learners, their probabilistic predictions are combined by the meta-learner via stacking. Large data preprocessing and augmentation improve generalization and robustness. Experimental results on a balanced MRI task achieve a test accuracy of 98.92%, as well as high precision, recall and F1-scores for all tumor classes are obtained. The confusion matrix and the ROC analyses demonstrated a high-performance discrimination power and low misclassification. The proposed method yields a practical and scalable solution for automated MRI-based brain tumor diagnosis, laying the foundation for future clinical decision-support systems.