Clinical neuroimaging cohorts are small. A pre-symptomatic Alzheimer’s study might have a few hundred subjects, which is nowhere near enough to train a deep network from scratch without overfitting. Transfer learning sidesteps this: pretrain a model on a large pooled dataset, then fine-tune it on the small clinical sample, so that the representations are learned where the data is abundant and only the task-specific mapping is learned where it is scarce.
The interesting assumption underneath it is scientific rather than technical. Pooling across conditions only helps if those conditions share underlying brain changes — which is exactly the hypothesis that motivates studying neurodegenerative and psychiatric conditions as a spectrum rather than in isolation.
CNS Projects¶
Hao-Ting Wang — pooling multiple datasets to pretrain a model, then using transfer learning to predict diagnoses within the smaller PREVENT-AD cohort. The premise is that Alzheimer’s disease, schizophrenia and autism spectrum disorder show significant overlap in underlying brain changes, so they are better investigated together
Getting started¶
Foundation models — the same idea at much larger scale
Deep learning — general context
Working with imbalanced data — small clinical cohorts usually bring a class imbalance problem with them