GeoSAE: Geometric Prior-Guided Sparse Autoencoder for Brain MRI Foundation Models
GeoSAE is a framework that utilizes a geometry-guided sparse autoencoder to tackle feature collapse within deep transformer layers of brain MRI foundation models. By leveraging the learned manifold structure of the model, it effectively prevents collapse and annotates features through age-deconfounded partial correlations. When applied to approximately 14,000 T1-weighted MRI scans sourced from the ADNI and AIBL datasets, GeoSAE successfully identifies a concise feature set that predicts the conversion from MCI to AD, achieving an AUC of 0.746 while utilizing just 2% of the embedding dimensions.
Key facts
- GeoSAE is a geometry-guided sparse autoencoder framework for brain MRI foundation models.
- It prevents feature collapse in deep transformer layers using the model's learned manifold structure.
- Features are annotated via age-deconfounded partial correlations.
- Applied to ~14k T1-weighted MRI scans from ADNI and AIBL datasets.
- Predicts MCI-to-AD conversion with AUC 0.746 using only 2% of embedding dimensions.
- Standard sparse autoencoders suffer from severe feature collapse in deep layers.
- Aging confounds clinical variables in Alzheimer's disease research.
- The framework identifies a compact, fully interpretable feature set.
Entities
Institutions
- Alzheimer's Disease Neuroimaging Initiative (ADNI)
- Australian Imaging Biomarkers and Lifestyle (AIBL)