An Explainable Deep Learning Approach for Brain Tumor Detection Using MobileNet and Grad-CAM Visualization

Brain tumor detection remains a significant challenge due to the complex variations in tumor appearance. Although deep learning models have demonstrated high accuracy, their limited interpretability hinders clinical adoption. To address this issue, this study integrates Gradient-weighted Class Activation Mapping (Grad-CAM) into Convolutional Neural Networks (CNNs) to enhance the visual interpretability of predictions. Grad-CAM extends Class Activation Mapping (CAM) and is applicable to a wide range of deep learning architectures. The primary contribution of this work is the demonstration that combining Grad-CAM with MobileNet architectures yields an interpretable and efficient framework for diagnosis of brain tumor, effectively balancing accuracy, computational efficiency, and clinical transparency. Using a Brain Tumor MRI dataset, MobileNetV4 achieved an accuracy of 98.29% with the shortest training time (1738.82 seconds) and an ROC accuracy of 99.96%. MobileNetV3 achieved 99.62% accuracy with an ROC accuracy of 99.92%. Grad-CAM effectively highlighted tumor regions while showing uniform attention in non-tumor cases, thereby reducing false positives. These results demonstrate that lightweight models can achieve a strong balance between predictive performance, training efficiency, and interpretability. The proposed framework thus supports the development of explainable and efficient diagnostic tools for clinical practice.