Multi-Class Brain Tumor Detection Using Transfer Learning and Interpretable Deep Models

Abstract

Accurate brain tumor detection remains critical yet challenging due to diagnostic complexity and variability in MRI interpretation. This study proposes a deep learning approach for automated multi-class brain tumor classification using transfer learning (TL), three pre-trained CNN models ResNet50, InceptionV3, and VGG16 were adapted and evaluated on a curated MRI dataset of 7,000+ images. Preprocessing, feature extraction, fine-tuning, and integration of Explainable AI (Grad-CAM, LIME, SHAP) ensured robust and interpretable results. ResNet50 achieved the highest performance with 98% accuracy, 0.92 F1-score, and 0.96 AUC, outperforming the other models across all metrics, with strong convergence and minimal misclassification. ResNet50’s architecture enabled deeper feature learning and improved generalization. Explainable AI visualizations confirmed model focus on tumor-relevant MRI regions, enhancing clinical interpretability. The findings position ResNet50 as an effective and explainable solution for MRI-based brain tumor classification, suitable for future real-world deployment and further expansion to mobile and multi-center applications.