A deep learning framework incorporating dynamic directed graph propagation achieved up to 89.21 percent diagnostic accuracy across large Alzheimer's and autism neuroimaging cohorts.