NYUMedML/CNN_design_for_AD

Implements CNN architecture optimizations (instance normalization, wide rather than deep networks, late spatial downsampling) validated on ADNI and external NACC cohorts, achieving 14% accuracy improvement over traditional volume/thickness models. Built with PyTorch and Clinica for preprocessing, the model directly processes structural MRI scans without requiring manual feature extraction, enabling both classification across cognitive stages and prediction of MCI-to-AD progression. Provides interpretable voxel-importance visualizations identifying disease-predictive imaging biomarkers across multiple brain regions.

174 stars. No commits in the last 6 months.