BRAIN-META: A reproducible CNN-vision transformer meta-ensemble pipeline for explainable brain tumor classification

This study presents BRAIN-META, a reproducible deep learning methodology designed for multi-class brain tumor classification using structural MRI. The proposed approach combines ten hybrid CNN-Vision Transformer (ViT) models with a meta-learning ensemble framework. The dataset includes 2D MRI images representing four tumor categories: glioma, meningioma, pituitary, and notumor. A standardized preprocessing pipeline involving image resizing, normalization, and CLAHE (Contrast Limited Adaptive Histogram Equalization) is applied to improve image quality and feature visibility. Ten pre-trained CNN architectures-DenseNet121, DenseNet169, DenseNet201, MobileNet, MobileNetV2, EfficientNetB0, EfficientNetB1, EfficientNetB4, InceptionV3, and Xception-are fused with Vision Transformer blocks to extract both local and global features. Each CNN-ViT model is trained independently, and the softmax outputs from validation data are used to generate stacked feature vectors. These vectors are input to two meta-learners, Logistic Regression and XGBoost, which are trained to produce final predictions. Evaluation metrics include accuracy, precision, recall, F1-score, and confusion matrix. XGBoost meta-learner achieved the highest accuracy of 97.10%, followed by Logistic Regression meta-learner at 97.03%, outperforming all individual base models. To enhance interpretability, Grad-CAM was employed, visually highlighting regions influencing classification. The proposed method is accurate, explainable, and modular, making it a strong candidate for clinical decision support in neuro-oncology.