Brain tumor detection using deep learning has become a focus of research due to its potential to improve patient diagnosis and management. This study presents a comparative analysis to evaluate how the fundamental design philosophies of convolutional neural network (CNN) architectures, namely sequential (VGG16), residual (ResNet50), and encoder-decoder (U-Net), influence performance on brain tumor classification tasks using a public MRI dataset. The data undergoes preprocessing, including normalization, data augmentation, and division into training and testing subsets. Evaluation is conducted using accuracy, precision, recall, and F1-Score metrics. The results show that ResNet50 consistently outperforms other architectures with an average accuracy of 95.35%, followed by VGG16 (93.93%). Conversely, U-Net, designed for segmentation, demonstrates the lowest performance (79%). The superiority of ResNet50 highlights the benefits of residual connections in addressing the vanishing gradient problem for classification tasks. The poor performance of U-Net confirms the hypothesis that architectural suitability for a specific task is more critical than complexity alone. This study highlights the strength of ResNet50 as a reliable approach for brain tumor classification. It underscores the importance of selecting an architecture specifically designed for the intended medical image analysis task. In addition, an analysis of model complexity and computational efficiency was conducted, showing that while ResNet50 achieved the highest accuracy, VGG16 provided a more favorable trade-off between performance and training time.