Comparative Performance of VGG16 and EfficientNetB0-Based Transfer Learning for Brain Tumor Classification

The classification of brain tumors using Magnetic Resonance Imaging (MRI) images is essential for early diagnosis but remains challenging due to tumor diversity. This study evaluates the effectiveness of two distinct architectural approaches for feature extraction: VGG16, representing a classic sequential design, and EfficientNetB0, a modern architecture optimized for parameter efficiency through compound scaling. Using a dataset of 2,870 MRI images categorized into four classes, we implemented a static transfer learning strategy by freezing all pre-trained ImageNet weights to act as fixed feature extractors. Features were extracted from specific layers, the final pooling layer for VGG16 and the Global Average Pooling (GAP) layer for EfficientNetB0. To optimize the discriminative power of these static features, this study implements an ensemble-fusion framework, specifically Voting and Stacking classifiers, to integrate the strengths of diverse base learners (SVM, Random Forest, and XGBoost) in identifying complex brain tumor patterns. Results demonstrate that VGG16-based models consistently and significantly outperform EfficientNetB0-based models across all evaluation metrics. The VGG16 + Random Forest combination achieved a peak accuracy of 73.9%, whereas EfficientNetB0-based models struggled, with peak accuracies reaching only 36.1%. Statistical analysis confirms that VGG16’s uniform convolutional layers provide more stable and discriminative feature representations for medical textures compared to the mobile-optimized blocks of EfficientNetB0 when weights are not fine-tuned. This research highlights that, for static feature extraction in resource-constrained environments, classic architectural stability offers superior reliability compared to modern parameter efficiency.