Hybrid Deep Learning Framework for Brain Tumor Classification Using MRI Images

Abstract

Brain tumor classification from magnetic resonance imaging (MRI) plays an important role in computer-aided diagnosis, as accurate identification of tumor categories can support clinical decision-making and reduce the burden associated with manual image interpretation. This study proposes a hybrid deep learning framework for multiclass brain tumor classification using MRI images. The proposed framework integrates two complementary transfer-learning architectures, ResNet50 and DenseNet201, to exploit their different feature-learning capabilities. Prior to classification, MRI images undergo a standardized preprocessing pipeline including resizing, noise reduction, contrast enhancement, and region-of-interest extraction to emphasize diagnostically relevant information. Data augmentation is applied exclusively to the training subset to improve model generalization while minimizing the risk of data leakage. ResNet50 and DenseNet201 are first evaluated independently under the same experimental conditions, followed by probability-level fusion of their classification outputs to construct the proposed hybrid model. The experimental results show that ResNet50 achieves an accuracy of 97.20%, while DenseNet201 achieves 98.10%. The proposed hybrid framework achieves the highest observed accuracy of 99.20%, demonstrating an improvement of 2.00 and 1.10 percentage points over ResNet50 and DenseNet201, respectively. In addition to quantitative evaluation, Grad-CAM is employed to provide visual interpretation of the model predictions by highlighting image regions contributing to the classification decision. Despite the promising results, further validation using independent datasets, patient-level data partitioning, repeated experiments, and statistical analysis is required to establish the generalizability and clinical reliability of the proposed approach. Overall, the findings demonstrate the potential of combining complementary transfer-learning models with preprocessing, ensemble fusion, and explainable artificial intelligence for automated brain tumor classification from MRI images.