CNNs vs. Hybrid Transformers for Brain Tumor Classification on the BRISC Dataset

Accurate and timely classification of brain tumors from Magnetic Resonance Imaging (MRI) is critical for effective treatment planning. The advent of deep learning has revolutionized medical image analysis, yet the performance of different model architectures is highly dependent on the quality of benchmark datasets and the specifics of the training methodology. This study presents a rigorous comparative analysis of four prominent deep learning architectures—ResNet18, EfficientNet-B0, MobileNetV3-Small, and the hybrid convolutional-transformer model MobileViTV2-100—for multi-class brain tumor classification. The models were trained and evaluated on the BRISC dataset, a large-scale, balanced collection of 6,000 T1-weighted contrast-enhanced MRI scans, comprising glioma, meningioma, pituitary, and no-tumor classes. Employing a 5-fold cross-validation protocol with a full fine-tuning strategy and robust regularization techniques, this study assesses models on both classification accuracy and computational efficiency. The results indicate that MobileViTV2-100, ResNet18, and EfficientNet-B0 achieve statistically comparable state-of-the-art performance, with mean test accuracies of 98.88%, 98.72%, and 98.72%, respectively. MobileNetV3-Small, while being the most parameter-efficient, demonstrated significantly lower accuracy at 96.94%. A key finding reveals a performance-efficiency paradox, where the largest model, ResNet18, exhibited the fastest inference latency (2.83 ms), challenging the conventional assumption that fewer parameters directly translate to greater speed. This comprehensive analysis underscores the strengths of hybrid architectures and provides critical insights into the practical trade-offs between model complexity, accuracy, and real-world deploy ability for clinical decision support systems.