Explainable Deep Learning for Brain Tumor MRI: Region-Level Validation of CNN and Transformer Architectures

Brain tumor diagnosis using Magnetic Resonance Imaging (MRI) is a critical task in medical imaging, where early and accurate detection plays a key role in treatment planning and patient outcomes. Although deep learning models have achieved promising results in tumor classification, their lack of transparency remains a major barrier to clinical adoption. In this study, we evaluated multiple deep learning architectures—including customized Convolutional Neural Networks (CNNs), VGG16, ResNet50, EfficientNetB0, and Vision Transformers (ViTs) for brain tumor classification. To address the interpretability gap, we introduce region-level validation as a post-training strategy that directly compares model predictions with expert medical annotations. This method enables us to determine whether models base their decisions on clinically meaningful tumor features rather than irrelevant patterns. Experiments on three publicly available brain tumor MRI datasets demonstrate that fine-tuned VGG16 and ViT models achieve both high classification accuracy and strong alignment with expert-annotated tumor regions. We argue that region-level validation provides evidence of trustworthy model behavior and represents an important step toward the safe integration of AI into radiology workflows.