3D CNN with Multi-ROI Features for Alzheimer's Classification from Brain MRI

As a leading neurodegenerative disorder, Alzheimer's Disease (AD) impacts millions globally and constitutes a critical public health concern. Early detection through Magnetic Resonance Imaging (MRI) analysis is crucial for patient management and therapeutic intervention. Existing deep learning approaches either use whole-brain volumes, which are computationally expensive, or focus on single regions like the hippocampus, potentially missing important multi-regional patterns. This study proposes a Multi-Region of Interest (Multi-ROI) 3D Convolutional Neural Network (CNN) with Attention Mechanism for AD classification. The proposed method extracts 12 Alzheimer-relevant brain regions simultaneously using the Automated Anatomical Labeling 3, version 1 (AAL3v1) atlas and processes them as multi-channel input to a custom five-stage residual 3D CNN with multi-level Convolutional Block Attention Module (CBAM). The selected regions cover memory network, default mode network, and semantic processing areas. Experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrate that the Multi-ROI approach achieves a mean Macro-AUC of 0.916 0.031 and Macro-F1 of 0.792 in three-class classification (Cognitively Normal, Mild Cognitive Impairment, and AD) with only 4.7 million parameters, representing an 86% reduction compared to standard whole-brain 3D ResNet-18 while capturing multi-regional patterns missed by single-ROI methods.