Efficient and Explainable 3D CNNs for Alzheimer's MRI Diagnosis

We present a configurable deep learning framework for classifying T1-weighted brain MRI scans into cognitively normal (CN), and Alzheimer’s disease (AD). The pipeline is optimized for high-performance computing (HPC) environments and integrates explainability methods for visual model interpretation.

Key contributions include a flexible, parameter-driven setup that enhances reproducibility, a memory-aware dataset chunking strategy for efficient GPU utilization, and built-in Grad-CAM visualizations for visual inspection of results. Using the ADNI dataset, our best binary classification (CN vs. AD) achieved 91.2% accuracy with a 3D ResNet10 model. The framework’s scalability, reproducibility, and explainability make it a robust tool for medical imaging research.
‍‍