Interpretable 2D Deep Learning for Alzheimer’s Detection from sMRI: A Lightweight Residual CNN Approach with Comprehensive Preprocessing and Stratified Data Partitioning

Neuroimaging is a promising modality for early AD detection, facilitating timely clinical intervention. This study proposes an enhanced deep learning framework that extracts critical AD biomarkers from structural MRI (sMRI) data acquired from the ADNI. Our novel CNN architecture integrates conventional convolutional layers with residual and skip connections for efficient feature extraction, achieving substantially lower computational cost than standard deep architectures such as VGG-16 (138 M), while remaining more parameter-intensive than highly compact architectures such as MobileNet and EfficientNet, which are designed explicitly for resource-constrained deployment. A comprehensive preprocessing pipeline converts 3D MRI scans into 2D slices through quality control (discarding slices with mean intensity < 5% of the maximum), bilinear resizing to 96 × 96 pixels, normalization using training-set statistics, and data augmentation. Stratified, subject-level data partitioning combined with robust statistical validation via bootstrapping demonstrates superior multiclass classification performance across AD, early and late MCI, and cognitively normal groups compared to state-of-the-art methods. Additionally, Grad-CAM-based interpretability maps were generated to highlight disease-relevant brain regions, confirming consistent activation around the hippocampus and temporal lobe.