Explainable Transfer Learning Framework for Multi-Class Brain Tumor Classification from MRI Images Using Comparative CNN Architectures
Accurate and explainable automated brain tumor classification using magnetic resonance imaging (MRI) remains an important challenge in medical image analysis. Transfer-learning studies often report high accuracy on public MRI benchmarks, but comparisons can be difficult to interpret when architectures are trained under different protocols or evaluated without statistical and explainability analyses. Here, we present an explainable transfer-learning framework for four-class brain tumor classification (glioma, meningioma, pituitary tumor, and no-tumor) in which MobileNetV2, ResNet50, and EfficientNetB0 are compared under a common training protocol. Models were trained and evaluated on the publicly available Brain Tumor MRI Dataset containing 7,200 T1-weighted contrast-enhanced axial images (5,600 training and 1,600 testing images, balanced across the four classes). The three backbones were fine-tuned using the same two-phase training strategy and compared using accuracy, precision, recall, F1-score, one-vs-rest ROC-AUC, confusion matrices, Cohen's kappa, Wilson 95% confidence intervals, McNemar paired tests, ablation analysis, and computational-cost measures. ResNet50 achieved the highest accuracy (96.06%; 95% CI: 94.99–96.91%) and macro-F1 (0.960), with a macro-averaged ROC-AUC of 0.990 and Cohen's kappa of 0.948. Its difference from EfficientNetB0 was statistically significant (McNemar, p = 0.003), whereas the difference from MobileNetV2 was not (p = 0.263). MobileNetV2 retained 99.6% of ResNet50's accuracy with substantially lower parameter and arithmetic cost. Per-class analysis identified glioma as the most difficult class. Grad-CAM was examined on correct and incorrect predictions and cross-checked with complementary attribution methods; a quantitative border-mass analysis over 200 correctly classified images per group did not support the specific hypothesis that no-tumor predictions were driven by a top-border watermark. Overall, the study supports controlled statistical, computational, and explainability analysis as a useful framework for benchmarking brain tumor MRI classifiers, while external validation and broader robustness assessment remain necessary before clinical use.