A controlled benchmark of CNN architectures for MRI-based brain tumor detection: Custom networks and transfer learning versus Vision Transformer baselines under a unified image-processing pipeline

Magnetic resonance imaging (MRI) is the reference modality for non-invasive brain-tumor assessment, but visual interpretation is slow, subject to inter-observer variability and hard to scale. Computer-aided systems based on convolutional neural networks (CNNs) can help, yet the biomedical image-processing literature lacks fair, head-to-head comparisons of custom and pre-trained models under a single pipeline. We present a controlled benchmark of twelve architectures — six custom CNNs, four pre-trained transfer-learning CNNs (VGG-16, VGG-19, ResNet-50 and MobileNetV2) and two Vision Transformers (ViT-B/16 and Swin-Tiny) — for binary tumor-versus-no-tumor classification in MRI, all trained under the same data, augmentation, splits, optimizer and learning-rate sweep, with model-appropriate input normalization. The Vision Transformer ViT-B/16 attains the best accuracy (0.9665) and F1 score (0.9662), whereas the hierarchical Swin-Tiny reaches only 0.9297 and does not beat the CNNs, indicating that the transformer advantage is architecture-specific rather than generic. With backbone-appropriate preprocessing the transfer-learning CNNs are strong (VGG-16 0.9521, ResNet-50 0.9473) and the best custom network trained from scratch (Arch-3, 0.9489) is competitive, so the gap between families is narrow; MobileNetV2 offers the best accuracy–efficiency trade-off. A preprocessing ablation shows that input normalization is a decisive, commonly neglected factor for frozen transfer learning: under a naive [ 0 , 1 ] pipeline ResNet-50 collapses to 0.8147 and Swin-Tiny to 0.8546, while MobileNetV2 and ViT-B/16 remain robust. Grad-CAM confirms that MobileNetV2 attends to tumor regions. The benchmark, released with public code, offers reproducible guidance for selecting architectures in clinical decision-support and edge-deployment scenarios.