A Stacked CNN Approach for Accurate Classification of AD Severity from T1-Weighted MRI Slices

This paper proposes a stacked convolutional neural network (CNN) architecture for classifying Alzheimer’s disease (AD) severity using T1-weighted MRI scans from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). 3D MRI volumes are preprocessed through intensity correction, spatial normalization to a standard template, and skull stripping, then sliced into informative 2D images based on anatomical landmarks and entropy measures. Data augmentation techniques further enhance the dataset while reducing overfitting. The stacked CNN, fine-tuned via transfer learning, extracts both local details and global structural features crucial for distinguishing healthy controls, mild cognitive impairment, and AD. A hybrid loss function combining cross-entropy and triplet loss improves the model’s discriminative power by clustering similar features and maximizing inter-class separation. Experimental results indicate high classification accuracy and robust performance, highlighting the potential of our approach for early AD diagnosis and severity assessment.