| name | geosae-brain-mri-sae |
| description | GeoSAE methodology for interpretable brain MRI foundation model annotation using geometry-guided sparse autoencoders with age-deconfounded partial correlations. Prevents SAE feature collapse in deep transformer layers, extracts biomarkers from frozen brain MRI foundation models. Achieves MCI-to-AD conversion prediction (AUC 0.746) with 2% embedding dimensions, cross-cohort replication (r=0.97). Use when: GeoSAE, brain MRI foundation model interpretability, sparse autoencoder for medical imaging, Alzheimer's biomarker discovery, age-deconfounded analysis, SAE feature collapse prevention, geometric prior SAE, brain MRI annotation, ADNI AIBL MRI analysis, Braak staging localization, MCI conversion prediction.
|
GeoSAE: Geometry-Guided SAE for Brain MRI Foundation Model Annotation
Nerrise et al., Stanford University, arXiv:2605.01829 (May 2026)
CVPR Workshop on Computer Vision for Clinical Applications (CV4Clinical) 2026
Problem
Brain MRI foundation models learn rich anatomical representations, but interpreting
what clinical information they encode remains difficult. Standard SAEs suffer from
severe feature collapse in deep transformer layers. In Alzheimer's research,
aging confounds nearly every clinical variable, making naive annotation unreliable.
Core Method
GeoSAE uses the foundation model's learned manifold geometry to prevent feature
collapse and annotates surviving features via age-deconfounded partial correlations.
Architecture
Brain MRI (T1-weighted)
→ Frozen Foundation Model (e.g., SynthSeg, FreeSurfer-style)
→ Layer-wise activations
→ GeoSAE (geometry-guided sparse autoencoder)
→ Interpretable features
→ Age-deconfounded partial correlation annotation
→ Clinical biomarker mapping
Key Innovations
- Geometric prior guidance: Uses the manifold structure learned by the foundation
model to guide SAE training, preventing feature collapse in deep layers
- Age-deconfounded annotation: Partial correlations control for age, isolating
disease-specific signals from normal aging effects
- Cross-cohort replication: Features replicate across ADNI → AIBL without
retraining (r=0.97)
Results
| Metric | Value |
|---|
| MCI-to-AD AUC | 0.746 |
| Embedding dimensions used | 2% |
| Cross-cohort replication | r=0.97 |
| Comorbidity-annotated features | Chance-level |
| Datasets | ~14k T1 scans (ADNI + AIBL) |
Key Findings
- Compact interpretable feature set predicts MCI-to-AD conversion using only 2%
of embedding dimensions
- Comorbidity-annotated features achieve only chance-level performance, suggesting
GeoSAE captures disease-specific rather than comorbid signals
- Neuroanatomical localization consistent with Braak staging of AD pathology
- Cross-cohort generalization without any retraining needed
Datasets
- ADNI: Alzheimer's Disease Neuroimaging Initiative
- AIBL: Australian Imaging Biomarkers and Lifestyle Study
- Total: ~14,000 T1-weighted MRI scans
Usage Patterns
1. Biomarker Discovery from Frozen Models
Apply GeoSAE to any frozen brain MRI foundation model to extract interpretable
clinical biomarkers without retraining the base model.
2. Age-Deconfounded Clinical Analysis
Use partial correlation annotation to separate disease effects from normal aging,
critical for neurodegenerative disease research.
3. Cross-Cohort Validation
Leverage geometry-guided features that replicate across different datasets without
retraining, enabling multi-site biomarker validation.
4. SAE Feature Collapse Prevention
Use geometric priors from the foundation model's manifold structure to guide SAE
training in deep transformer layers.
Limitations
- Requires a pretrained brain MRI foundation model
- T1-weighted MRI only (no multi-modal extension shown)
- Age deconfounding assumes linear age effects
- Focused on AD/MCI — extension to other diseases needs validation
Code
Related Work
- Sparse Autoencoders (SAEs) for LLM interpretability
- Brain MRI foundation models (SynthSeg, SynSegHD, etc.)
- Alzheimer's disease neuroimaging biomarkers
- Braak staging of AD pathology
- Age-deconfounded neuroimaging analysis
