Multi-Orientation Hippocampus-Centered 3D CNN with Attention Mechanism for Alzheimer's Disease Classification from MRI Scans
turanli, m.
Show abstract
Alzheimers disease detection faces challenges in capturing hippocampal atrophy across multiple anatomical orientations. This study presents a multi-orientation hippocampus-centered 3D CNN with attention mechanism for automated classification. The architecture processes three parallel 40x128x128x1 volumes from sagittal, axial, and coronal orientations. Each branch employs Conv3D layers with dilated convolutions and attention-based feature fusion. Training on ADNI dataset (1008 subjects: 652 normal, 356 Alzheimers) using focal loss achieves AUC-PR values of 0.982-0.990 across five-fold cross-validation. The hippocampus-centered preprocessing uses MNI152 registration and FIRST segmentation. Results demonstrate superior performance with interpretable attention weights for clinical deployment.
Matching journals
The top 9 journals account for 50% of the predicted probability mass.