Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Jul 24·5 min read
DS@GT ARC at ImageCLEFmed GANs 2026: Geometric Filtering for Privacy-Preserving CT Slice Generation
TL;DR
A privacy-preserving framework for generating synthetic lung CT slices was developed, achieving a strong balance between realism and privacy. The best model scored 0.549 on Privacy Preservation and 0.3290 on visual fidelity metrics.
✦ Why It Matters
Consider implementing geometric filtering techniques to enhance privacy in your synthetic medical imaging projects.
Key Takeaways
- Explore Optimal Transport Conditional Flow Matching for privacy-oriented training.
- Utilize autoencoder embeddings for effective image filtering.
- Implement Determinantal Point Processes to reduce memorization risks.
- Address deeper anatomical identity issues in synthetic image generation.
How It Works
The framework employs Optimal Transport Conditional Flow Matching to generate synthetic CT slices while integrating a 'Supervisor' pipeline that filters outputs. This filtering process uses learned geometric latent spaces, leveraging autoencoder embeddings to ensure generated images are both realistic and privacy-preserving.
Determinantal Point Processes help in selecting diverse candidates, while Stein Kernel Thinning further refines the output, reducing the risk of revealing patient-specific information.
⚠ The Catch
Despite advancements in privacy preservation, the framework still faces challenges with persistent patient re-identification, indicating that preventing direct image copying alone is insufficient to protect deeper anatomical identities.