Geometric Filtering Framework Achieves 0.549 Privacy Score on Synthetic CTs

A team from DS@GT ARC (Eric Regina, Richard Arnaud, Samir Hadi Cisneros) has submitted a paper to the ImageCLEFmed GANs 2026 challenge presenting a new privacy-preserving pipeline for synthetic lung CT slice generation. The core innovation is a three-stage approach: first, they use Optimal Transport Conditional Flow Matching (a generative model that transports noise to realistic CT distributions with conditional constraints) trained with privacy-oriented regularization. Second, they apply a 'Supervisor' post-generation filter that evaluates candidates in learned geometric latent spaces using autoencoder embeddings, Determinantal Point Processes (to ensure diversity), and Stein Kernel Thinning (to select representative samples). This reduces the risk of directly copying real patients' CT slices.

Official challenge results show a strong realism–privacy trade-off: the best configuration achieved a Privacy Preservation Score of 0.549 (higher is better) while maintaining competitive visual fidelity with an FID of 0.3290 (lower is better). However, the authors note that while geometric filtering reduces nearest-neighbor memorization and membership-inference leakage, patient re-identification scores remained persistently high. This indicates that preventing direct image copying is insufficient to erase deeper patient-specific anatomical identity (e.g., lung shape, lesion patterns). The work highlights an important frontier: future methods must tackle not just pixel-level memorization but also higher-order structural signatures unique to individuals.