Explainable Transfer Learning for Brain Tumor MRI Classification: Leakage-Aware Evaluation, Grey Wolf Optimizer-Based Hyperparameter Tuning, and Quantitative Attribution Analysis

Background/Objectives: Accurate MRI-based brain tumor classification requires leakage-free performance estimates and inspectable decisions. We compared pretrained convolutional neural networks (CNNs), evaluated Grey Wolf Optimizer (GWO)-based hyperparameter tuning against alternative search strategies, and quantified Grad-CAM explanations. Methods: Seven ImageNet-pretrained CNNs were compared with duplicate-aware 10-fold cross-validation on 5966 images of a public four-class dataset; 1057 images were held out. GWO tuning of VGG16 was compared with five other search strategies at an equal budget. Ablation, cost-sensitive training, external validation (BDNeuro-MRI), and Grad-CAM analyses of all test images were performed. Results: VGG16 had the highest cross-validation accuracy under the original protocol (96.73 ± 1.24%) but not under full fine-tuning. GWO-VGG16 reached 98.68% hold-out accuracy (95% CI 97.92–99.34%). Because 39.7% of test images had a duplicate in the development data, accuracy was also assessed on 637 duplicate-free test images (98.27%; 95% CI 96.93–99.03%) and on 1888 non-overlapping external images (87.29%; 95% CI 85.75–88.77%). GWO was not superior to alternative search strategies. Grad-CAM maps showed little border attention and were model-dependent, but lesion localization remained unverified. Conclusions: High internal accuracy did not fully transfer to external data; without patient-level validation and expert annotations, the framework remains a research-stage tool, not a validated diagnostic system.