Comparative Evaluation of ResNet50 and MobileNetV2 Enhanced with Explainable AI for Brain Tumor MRI Classification

Abstract

Accurate identification of brain malignancies via MRI is a cornerstone of effective neuro-oncological intervention. While deep learning offers high-speed automated analysis, the unclear reasoning mechanisms of these algorithms often limit their adoption in clinical environments. This study presented a comparative evaluation of two distinct convolutional neural network (CNN) architectures—ResNet50 and MobileNetV2—to determine the optimal balance between classification depth and computational efficiency. Both models were evaluated on an MRI dataset to distinguish between healthy and tumor-containing scans, with class weighting applied to mitigate the effects of class imbalance. Experimental results demonstrate high performance across both architectures. ResNet50 achieved an accuracy of 95.47%, supported by a precision of 95.16%, a recall of 95.44%, and an F1-score of 95.29%. In comparison, the lightweight MobileNetV2 reached an accuracy of 94.80%, with precision, recall, and F1-score recorded at 94.33%, 95.00%, and 94.63%, respectively. To address the requirement for medical transparency, Grad-CAM was integrated to generate visual evidence for each prediction. These heatmaps successfully localized tumor regions, aligning internal model logic with radiological findings. By benchmarking a high-capacity residual network against an efficient, mobile-ready architecture, this study confirms that both frameworks, when paired with explainability tools, provide a reliable secondary screening layer. This integration of high-performance classification with visual interpretability offers a clear path toward deploying trustworthy diagnostic tools in clinical workflows.