Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a novel privacy-preserving framework for generating synthetic lung CT slices, specifically designed for the Image-CLEFmed GANs 2026 challenge. By integrating Optimal Transport Conditional Flow Matching with a post-generation filtering mechanism that utilizes autoencoder embeddings and advanced statistical techniques, the model achieves a notable balance between realism and privacy, evidenced by a Privacy Preservation Score of 0.549 and an FID of 0.3290. However, despite these advancements, the study reveals that while geometric filtering mitigates certain privacy risks, it does not fully eliminate the potential for patient re-identification, underscoring a critical area for further research.
Achieving a Privacy Preservation Score of 0.549, this framework reveals that even advanced filtering techniques can't fully safeguard against patient re-identification in synthetic medical images.
We present a privacy-preserving framework for synthetic lung CT slice generation developed for the Image-CLEFmed GANs 2026 challenge. The approach combines Optimal Transport Conditional Flow Matching with privacy-oriented training and a post-generation "Supervisor" pipeline that filters generated candidates in learned geometric latent spaces using autoencoder embeddings, Determinantal Point Processes, and Stein Kernel Thinning. Official results show a strong realism-privacy trade-off, with the best-performing model achieving a Privacy Preservation Score of 0.549 and competitive visual fidelity with an FID of 0.3290. While the proposed geometric filtering substantially reduces nearest-neighbor memorization and membership-inference leakage, persistent patient re-identification scores indicate that preventing direct image copying is not sufficient to remove deeper patient-specific anatomical identity, highlighting an important frontier for future privacy-preserving medical image generation.