NERVE: Neuroimaging Embedding Representation via Variational Encoding

Motivation: Inspired by large language models, we aim to apply embedding techniques to tackle high-dimensional MRI data and limited labels, aiming to improve Alzheimer's disease classification. Goal(s): Develop an open, general-purpose brain MRI embedding model using variational autoencoders to create informative embeddings for downstream neuroimaging tasks. Approach: Collected large MRI datasets, extracted hippocampus and amygdala patches, trained VAEs to generate embeddings (NERVE), and used these embeddings to train a ResNet classifier for AD detection. Results: NERVE embeddings significantly improved AD classification, with the open dataset model achieving an F1-score of 0.92, outperforming models using raw or preprocessed MRI data. Impact: NERVE encodes brain MRI into a compressed embed. A generalized and open-source NERVE model offers broad applications in neuroscience.