Brain Tumor Detection and Classification from MRI Images with Systematic Screening of a CNN–Transformer Hybrid Framework

Abstract

Brain tumors are the most fatal diseases of the central nervous system. Accurate, early MRI classification can support timely clinical decisions and treatment planning. This study systematically identifies an effective and efficient hybrid architecture combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for multi-class brain tumor classification from 2D MRI images through a systematic screening evaluation. In this study, 12 CNN–Transformer pairs were evaluated under an identical training procedure using the original Kaggle train/test split. A late-concatenation fusion of DenseNet-121 and Swin-Tiny yielded the highest Kaggle testing accuracy of 94.38% during the screening stage. To provide a more reliable and fair comparison, a 10-fold stratified Cross-Validation (CV) was conducted on the Kaggle Training set only. The best-ranked pair was EfficientNet-B0 + Swin-Tiny, achieving a mean CV accuracy of 98.29% ± 0.55% with selective fine-tuning. The final model was trained on the Kaggle Training set, evaluated on the Kaggle Testing set, and tested on an independent external dataset (TEST2). The final model achieved 93.94% accuracy on the Kaggle Test Set and 87.21% accuracy on TEST2. The result indicates a notable domain shift under external validation. In addition, paired statistical significance tests among top model pairs, ablation studies on the fusion strategy and trainable ratio, and computational cost analysis (training time, inference latency, GPU memory, and parameter counts) are reported.