Explainable artificial intelligence for brain tumor classification via fine-tuned transfer learning

Brain tumors remain one of the most life-threatening diseases worldwide, marked by the abnormal and aggressive proliferation of cells within the brain. Accurate and early detection is critical to improving treatment outcomes and reducing patient mortality rates. However, manual analysis of brain MRI scans is often time-consuming, prone to inter-observer variability, and lacks scalability. To address these challenges, this study proposes an efficient and interpretable deep learning (DL) model for brain tumor classification, utilizing transfer learning with the ResNet50 architecture. The model is trained to distinguish among three tumor types—glioma, Meningioma, and pituitary tumors, and normal brain MRI scans, using a fine-tuned network combined with extensive data augmentation strategies to mitigate overfitting and enhance generalization, particularly on limited-size medical datasets. The proposed model was trained and evaluated on three publicly available multiclass brain tumor MRI datasets comprising multi-planar MRI scans (axial, coronal, and sagittal) collected from patients with diverse ages, tumor grades, and demographic backgrounds. In particular, the BT-large-4c dataset from Kaggle includes 3264 MRI scans of brain tumors and healthy controls across four classes (395 normal, 827 pituitary, 826 glioma, and 822 Meningioma). The proposed system achieved outstanding performance, attaining 99.41% accuracy, 99.15% precision, 99.09% sensitivity, 98.18% specificity, and a 98.91% F1-score on the test dataset. Unlike conventional black-box models, this work emphasizes interpretability by integrating SHAP (Shapley Additive explanations), an Explainable AI (XAI) technique that visualizes and quantifies feature importance at the pixel level. This transparency facilitates greater clinical trust and understanding by elucidating the rationale behind model predictions. Moreover, the proposed ResNet50-based model consistently outperformed other established transfer learning architectures, including VGG16, MobileNet, and DenseNet201, across key evaluation metrics. By combining high diagnostic performance with SHAP-based interpretability, the system provides transparent and trustworthy predictions that can meaningfully support radiologists and healthcare professionals in routine clinical practice and the early diagnosis of brain tumors.