What question did this study set out to answer?

To develop and evaluate a framework that combines transfer and ensemble learning for classifying brain tumors using MRI images.

March 17, 2026Open Access

A deep transfer learning and ensemble learning framework for brain tumor detection and classification in MRI images

Key Points

To develop and evaluate a framework that combines transfer and ensemble learning for classifying brain tumors using MRI images.
Integrated transfer learning and ensemble learning for tumor classification.
Analyzed and compared multiple deep learning architectures: VGG, Inception, ResNet, DenseNet.
Selected six convolutional neural networks for the classification task.
Applied a weighted voting strategy in the ensemble of top-performing models.
Utilized Grad-CAM for explainability of the model's decisions.
Achieved 99.28% accuracy, 98.62% precision, 99.56% recall, and 98.58% F1-score on the BT-7023 dataset.
Outperformed existing models with 98.80% accuracy on the cross-dataset validation.
Demonstrated a 5%–13% performance improvement over previous methods.
Enhanced model transparency using XAI techniques to highlight key regions in MRI images.

Abstract

This study proposes a novel framework that integrates Transfer Learning (TL) and Ensemble Learning (EL) for multi-class classification of brain tumors using Magnetic Resonance Imaging (MRI) images. The classification targets four categories: glioma, meningioma, pituitary tumor, and no tumor. Several state-of-the-art deep learning architectures from the VGG, Inception, ResNet, and DenseNet families were analyzed and compared. Based on their performance, six Convolutional Neural Networks (CNNs)—VGG16, VGG19, InceptionV3, Xception, ResNet152, and DenseNet201—were selected for transfer learning to perform the classification task. To enhance classification accuracy, ensemble learning was applied using a weighted voting strategy, where the accuracy of each model served as its weight. The four top-performing models VGG16, Xception, ResNet152, and DenseNet201 were combined in this ensemble. To further interpret the model’s decisions, Explainable Artificial Intelligence (XAI) technique, specifically Grad-CAM, was employed to highlight key regions in the MRI images that influenced the classification outcomes, thereby improving model transparency. The models were trained and evaluated on the publicly available BT-7023 dataset from Kaggle, consisting of 7,023 MRI images (5,712 for training and 1,311 for testing). Performance was assessed using accuracy, precision, recall, and F1-score. The proposed ensemble method achieved excellent results, with 99.28% accuracy, 98.62% precision, 99.56% recall, and a 98.58% F1-score. To verify the robustness of the approach, cross-dataset validation was conducted using an external dataset from Figshare. The proposed ensemble outperformed existing models on this dataset, achieving 98.80% accuracy. Overall, the ensemble framework demonstrated a 5%–13% performance improvement over existing methods, underscoring its effectiveness and reliability for early detection and classification of brain tumors. Additionally, the XAI-based analysis confirmed the model’s focus on clinically relevant tumor regions, enhancing interpretability and trustworthiness in medical applications.

Ask AI

Helpful

Bookmark

View Full Paper