Explainable Diagnostic Framework for Brain Tumor Detection: Integrating Decision Optimization and Grad-CAM Interpretability

Background: The integration of Deep Learning (DL) in neuro-oncology has significantly improved diagnostic accuracy; however, the "black-box" nature of these models remains a fundamental barrier to clinical trust and regulatory approval. This study addresses the critical need for interpretability in medical AI by developing a transparent diagnostic framework. Methods: Building upon our previous architectural innovations in two-stage transfer learning and asymmetric class weighting, we propose a fine-tuned MobileNetV2-based system enhanced with Gradient-weighted Class Activation Mapping (Grad-CAM). The model incorporates decision threshold optimization (0.3 instead of 0.5) to eliminate False Negatives (FN), a strategy validated in our earlier research. We evaluate the model's spatial reasoning on a diverse test set of 10 unseen MRI scans, comprising both pathological and healthy samples. Results: The system achieved a remarkable training accuracy of 100% and a validation accuracy of 99.51% (𝐿𝑜𝑠𝑠 = 0.0185). By implementing the optimized threshold, we secured a 100% recall for the tumor class. Grad-CAM visualizations demonstrated a high degree of spatial fidelity, where the model consistently localized hyper-intense necrotic and enhancing regions of tumors with high precision. Conclusion: The proposed framework offers a dual-validation mechanism: superior statistical performance and visual accountability. This transparency allows clinicians to verify the biological relevance of the AI's focus, bridging the gap between computational power and clinical diagnostic safety.