To develop and evaluate a deep learning model using brain MRI for automated staging of dementia in Alzheimer’s disease, including the diagnostically challenging very mild stage.
Alzheimer’s disease (AD) is a progressive neurodegenerative disorder where early detection is critical, as timely interventions can slow progression. Traditional cognitive tools like the MMSE have limited sensitivity to early decline, while MRI offers objective insights into structural brain changes preceding symptoms.
A publicly available brain MRI dataset comprising 33,984 augmented images was categorized into four groups: Non-Demented (9,600), Very Mild Demented (8,960), Mild Demented (8,960), and Moderate Demented (6,464). The data were split into training (50%), validation (20%), and testing (30%) sets. Transfer learning with MobileNet architecture was implemented. Model performance on the held-out test set was assessed using accuracy, precision, recall, F1-score, and confusion matrix. Grad-CAM heatmaps were generated for interpretability.
On the test set (n = 10,196), the model achieved an overall accuracy of 98.2% and test loss of 0.057. Macro-averaged precision, recall, and F1-score were all 0.98. Class-wise F1-scores were: Mild Demented = 0.99, Moderate Demented = 1.00, Non-Demented = 0.97, and Very Mild Demented = 0.97. Confusion matrix analysis showed minimal misclassification. Grad-CAM overlays consistently emphasized hippocampal and temporal regions, correlating with known AD pathology.
Deep learning models applied to structural MRI can accurately classify multiple AD stages, including the very mild stage often missed by conventional tools. These findings support the potential of MRI-based deep learning as a non-invasive, objective approach for early diagnosis and staging of Alzheimer’s disease.