Brain Tumor Classification in MRI: Insights From LIME and Grad-CAM Explainable AI Techniques

Artificial intelligence (AI) has become increasingly vital in brain tumour diagnosis using magnetic resonance imaging (MRI). This study serves as a reproduction and interpretability case study, evaluating the behaviour of established convolutional neural network (CNN) architectures: VGG16, ResNet50, and EfficientNet through the implementation of Explainable AI (XAI) to interpret their black-box predictions of tumour classes. All models were trained with early stopping, achieving high classification accuracy: VGG16 and ResNet50 exceeded 98%, while EfficientNetB0 achieved 96%; however, misclassification between Glioma and Meningioma remains a challenge, suggesting overlapping features. VGG16 exhibited more stable validation accuracy; ResNet50 demonstrated fluctuations due to its deeper architecture and broader feature activations, while EfficientNet maintained a more gradual learning curve, attributed to its compound scaling. To enhance interpretability, Local Interpretable Model-agnostic Explanations (LIME) and Gradient-Weighted Class Activation Mapping (Grad-CAM) were applied. LIME segments images into superpixels and analyses their contributions through repeated perturbations. LIME revealed that CNNs sometimes relied on irrelevant background features, including brain edges and anatomical distractions, leading to misclassification. Meanwhile, Grad-CAM, which highlights class-discriminative regions through gradient-based heatmaps, demonstrated key architectural differences: ResNet50 activated broader spatial areas, VGG16 focused on more localized tumour features, while EfficientNet often considered wider anatomical contexts that extended to structures such as eye sockets and the nasal cavity. These results underscore how architectural depth and feature aggregation influence classification performance and reveal potential limitations in CNN-based classification. This study contributes a comparative interpretability perspective, demonstrating how architecture-specific behaviours influence feature attribution and reliability in CNN-based brain tumour classification.