<![CDATA[Brain tumors are life-threatening central nervous system disorders requiring early and accurate diagnosis for effective clinical management. Although MRI is the standard modality for detection, manual interpretation remains prone to inconsistency, particularly for complex cases such as glioma. This study proposes an explainable deep learning framework integrating EfficientNet-B2 with a threshold-based two-stage classification scheme and Grad-CAM interpretability analysis. In the first stage, a one-versus-rest binary classifier with an optimized threshold (τ = 0.20) performs glioma detection; the second stage classifies remaining cases into meningioma, pituitary tumor, or normal. The dataset comprises 7,023 MRI images across four classes from a public Kaggle repository. Preprocessing includes CLAHE contrast enhancement, normalization, and augmentation. EfficientNet-B0 serves as the baseline. EfficientNet-B2 achieves 97.9% overall accuracy, outperforming the baseline (96.7%), with a glioma F1-score of 0.988 at the optimal threshold. Grad-CAM visualizations confirm the model focuses on anatomically relevant regions, enhancing transparency and clinical trustworthiness. The proposed framework demonstrates that combining architectural capacity, threshold-based inference, and explainability yields a reliable system for computer-aided brain tumor diagnosis.]]>
Copyrights © 2026
