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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate brain tumor classification is crucial for effective treatment planning.
• Existing methods may lack robustness or interpretability in multi-class scenarios.

Purpose of the Study:

• To develop and evaluate BRAIN-META, a reproducible deep learning methodology for multi-class brain tumor classification using structural MRI.
• To combine Convolutional Neural Network (CNN) and Vision Transformer (ViT) models with meta-learning for enhanced classification performance.

Main Methods:

• A hybrid CNN-ViT architecture was developed, integrating ten pre-trained CNNs with ViT blocks.
• A meta-learning ensemble framework using Logistic Regression and XGBoost was employed for final predictions.
• Standardized preprocessing and Grad-CAM for interpretability were utilized.

Main Results:

• The XGBoost meta-learner achieved a top accuracy of 97.10%, outperforming individual base models.
• Logistic Regression meta-learner reached 97.03% accuracy.
• Grad-CAM visualization confirmed model interpretability by highlighting relevant image regions.

Conclusions:

• BRAIN-META demonstrates high accuracy and explainability for brain tumor classification.
• The modular and reproducible methodology shows potential for clinical decision support in neuro-oncology.