OPTIMIZING BRAIN TUMOR CLASSIFICATION WITH FREEZE-5 VGG16 AND DATASET FUSION

Magnetic resonance imaging (MRI)-based brain tumor classification is pivotal for early diagnosis and treatment planning. This study enhances the VGG16 pretrained model through freeze-5 fine-tuning (i.e., freezing the first five convolutional layers) and dataset fusion of two public repositories, yielding 5,023 training and 1,311 testing images. Preprocessing includes normalization and grayscale-to-RGB conversion, followed by moderate augmentation (rotation ≤ 15°, shift ≤ 0.1, zoom ≤ 0.1, brightness [0.9–1.1]). The base VGG16 (without top layers) is extended with GlobalAveragePooling2D, Dense (1024, ReLU), Dropout (0.5), and Dense (4, softmax) layers. The model is compiled with the Adam optimizer (lr=1e-4), EarlyStopping, and ReduceLROnPlateau callbacks. On the test set, the proposed configuration achieves peak accuracy of 99.16 % and macro-F1 of 0.99, outperforming prior hybrid approaches. An ablation study confirms that the freeze-5 strategy combined with data augmentation significantly boosts generalization without overfitting. These results underscore the critical role of optimal layer-freezing and dataset fusion in brain tumor classification. Future work will explore ensemble architecture and real-time clinical deployment.