Empirical Evaluation of Post Hoc Interpretability Methods for Clinical Trust in Deep Learning Diagnostics

Background/Objectives: Brain tumor classification from magnetic resonance imaging (MRI) has been significantly improved through deep learning techniques; however, the lack of interpretability remains a major challenge for clinical adoption. Methods: This study proposes an explainable artificial intelligence (XAI)-driven framework for multiclass brain tumor classification using a baseline convolutional neural network (CNN) and four transfer learning models, namely ResNet50, DenseNet121, MobileNetV2, and EfficientNetB0. The models were trained and evaluated on a publicly available brain MRI dataset consisting of glioma, meningioma, and pituitary tumor classes. Their classification performance was assessed using accuracy, precision, recall, F1-score, receiver operating characteristic area under the curve (ROC-AUC), and confusion matrix analysis. To enhance transparency and interpretability, Grad-CAM, LIME, and SHAP were used to visualize and quantify the regions contributing to the classification decisions. Results: Experimental results state that the transfer learning models consistently outperformed the CNN by learning more discriminative feature representations while producing more localized and interpretable explanations. The agreement among Grad-CAM, LIME, and SHAP confirms that the proposed models predominantly relied on tumor regions rather than irrelevant background information, thereby improving the interpretability and reliability of the diagnostic predictions. Conclusions: The proposed framework provides an effective and explainable computer-aided diagnosis system that supports trustworthy deep learning-based brain tumor classification and has the potential for decision-making.