Early Detection of Mild Cognitive Impairment from Structural MRI Using Deep Learning and Explainable AI

Abstract

Alzheimer's disease is a progressive neurodegenerative disease that leads to memory loss, cognitive decline and brain atrophy. Mild Cognitive Impairment is the early transitional stage between normal ageing and Alzheimer's disease in which brain structural changes start to appear but daily life functioning is still preserved. Early and accurate detection of MCI from brain MRI is clinically hard due to the subtle signal, about 1-3% per year hippocampal volume loss and cortical thinning, which cannot be reliably identified by an human expert. In this work, we propose a deep learning pipeline to first apply HD-BET to skull strip T1-weighted 3D MRI volumes from Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset and retain only brain tissue information, then crop out 2D slices in the central brain region in 3 anatomical planes, and further train a fine-tuned ResNet50 on these slices for binary classification of cognitively normal (CN) subjects vs MCI subjects. The system reaches balanced accuracy of $\mathbf{8 3. 3 5 \%}$, AUC-ROC of $\mathbf{9 1. 3 0 \%}$. Gradientweighted Class Activation Mapping visualisations show that model attention spatially aligns with hippocampal and temporal lobe regions, which are the main anatomical biomarkers of early MCI, making the predictions interpretable by humans.