Explainable Deep Learning Model for Multi-Class Dementia Classification Using Brain Magnetic Resonance Imaging from Neurologist and Radiologist Perspectives

Abstract

Introduction: Although deep learning classifiers are often not clinically interpretable, magnetic resonance imaging (MRI) is central to the diagnosis of dementia. In the present work, a MobileNetV2 based multi-class dementia classification model is proposed. It is also supported by an anatomical validity evaluation of its Gradient-weighted Class Activation Mapping (Grad-CAM) explanations. This research was targeted to contribute to SDG 3 (good health and well-being) through a transparent and explainable AI development particularly for cognitive impairment disease. Methods: A transfer learning framework with inverse-frequency class weighting was applied to 86,437 MRI slices from the OASIS-1 dataset to cover four category of disease severity. An independent test set was applied to assess the performance. Subsequently, heatmap concordance evaluation was appraised qualitatively by a multidisciplinary panel of neurologists and radiologists. Results: The structure achieved an overall test accuracy of 81.95% and a test AUC of 0.9655. However, a low performance was found in the prodromal stages, witnessed by lower F1-scores for Mild (0.57) and Very Mild Dementia (0.58). According to reasonability audit that identified significant discrepancies, the heatmaps indicates partial anatomical concordance in majority cases, with only moderate dementia cases was reported as complete concordance. Conclusion: lightweight transfer learning exhibits high diagnostic accuracy. However, the discrepancy between predictive metrics and partial anatomical validity needs strict validation and leakage-free external evaluation before clinical implementation.