IMPROVED HYBRID DEEP LEARNING TECHNIQUE FOR BRAIN TUMOR CLASSIFICATION
Keywords:
Magnetic Resonance Imaging (MRI), VGG16, Convolutional Neural Networks (CNNs), Grad-CAMAbstract
Brain tumors present significant diagnostic challenges due to their visual variability and the limitations of manual MRI interpretation. This study proposes an interpretable ensemble deep learning model integrating VGG16, EfficientNet, LeNet, AltNet, and Vision Transformer (ViT) for multi-class brain tumor classification. Unlike the benchmark study which combined only CNN and VGG16 and achieved 98.41% accuracy with lower recall (91.4%) and F1-score (91.54%), our ensemble achieves 98.35% accuracy, 98.50% precision, 98.40% recall, and 98.45% F1-score. While accuracy remains comparable, the substantial improvements in recall and F1-score demonstrate superior clinical reliability. The integration of Gradient-weighted Class Activation Mapping (Grad-CAM) provides interpretable visual explanations of predictions a feature absent in the benchmark. Cross-dataset validation on the Figshare dataset confirms generalizability with 97.8% F1-score. This work contributes a more balanced, clinically relevant, and explainable solution for MRI-based brain tumor classification.
