Brain tumors are among the most critical neurological conditions, requiring accurate and early diagnosis to improve patient outcomes. Manual interpretation of MRI images by radiologists is time-consuming and prone to variability. This study proposes a deep learning approach based on transfer learning with the VGG16 architecture to classify brain MRI images into four categories: glioma, meningioma, pituitary tumor, and no tumor. The dataset consisted of 5,600 training images and 1,600 independent testing images, evenly distributed across the four classes. Training was conducted in two phases: first training a classifier head with the VGG16 backbone frozen, followed by fine-tuning the last eight layers with a substantially reduced learning rate, incorporating batch normalization and class-weighted balancing. The proposed model achieved a test accuracy of 93.0% with a macro-average precision, recall, and F1-score of 0.93, 0.93, and 0.93, respectively, evaluated on completely held-out test data. Error analysis revealed that glioma had the lowest recall (82%) among the four classes, predominantly misclassified as meningioma, while a secondary confusion pattern was observed between meningioma and pituitary tumors. These results demonstrate that a two-phase transfer learning strategy with VGG16 is effective for automated brain tumor classification from MRI images, while highlighting inter-class visual similarity as a key challenge for further improvement. A direct comparative evaluation against ResNet50 and MobileNetV2 under an identical training pipeline further confirmed that VGG16 achieved the best balance between classification accuracy, glioma recall, and computational efficiency among the three architectures.Unlike prior studies that primarily reported aggregate accuracy, this study provides a systematic error analysis linking glioma misclassification to MRI slice orientation variability, offering practical guidance for dataset curation in future brain tumor classification research.