Research map / Label-efficient and robust learning / Recognition and long-tail
Medical Manifestation-Aware De-Identification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Releases MeMa, a 42k-image synthetic patient-face dataset generated by fine-tuned Stable Diffusion from real eye-clinic photos, and proposes MedSem-DeID, a reversible medical-semantics-preserving face de-identification baseline.
Problem
Face de-identification for medical scenes is under-researched because large-scale patient face datasets are unavailable and common DeID methods destroy disease manifestations needed for diagnosis.
Approach
Generate privacy-safe patient faces by fine-tuning Stable Diffusion on real patient photos with disease-type prompts, injecting public FFHQ face attributes via IP-Adapter to reduce identity leakage, and use a diffusion-derived medical semantics encoder to preserve disease signs in a reversible DeID network.
Key results
- MedSem-DeID achieves 86.70% overall disease classification accuracy on MeMa validation, outperforming the second-best Disguise (76.01%).
- On MeMa-Seg tumor segmentation, the method reaches Dice 0.6775 and Jaccard 0.5453, the best among compared methods.
- On real-world Real-ECXHCSU cohort, Cohen's Kappa exceeds 0.82 across five disease categories, versus 0.0566 for digital mask on BCC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MedSem-DeID instead of Disguise | Medical-scene de-identification · MeMa validation | classification accuracy (All): 76.01 -> 86.70% | Helped |
| Use MedSem-DeID instead of Password | Medical-scene de-identification · MeMa-Seg validation | Dice: 0.6336 -> 0.6775 | Helped |
| Inject public face attributes during patient face synthesis (ablation) | Patient face generator · Generated patient faces | average identity leakage percentage: 71.83% -> 1.27% | Helped |
| Guide synthesis with real patient disease distributions (ablation) | Patient face generator · BCC generation | Wasserstein distance (disease): 0.256 -> 0.003 | Helped |
| Replace FFHQ training data with MeMa (ablation) | MedSem-DeID without medical priors · MeMa validation | disease classification accuracy: 42.13% -> 46.82% | Helped |
| Add medical feature f_med (ablation) | MedSem-DeID · MeMa validation | disease classification accuracy: 46.82% -> 69.94% | Helped |
| Add medical loss L_med (ablation) | MedSem-DeID · MeMa validation | disease classification accuracy: 46.82% -> 71.35% | Helped |
| Combine medical feature f_med and medical loss L_med (ablation) | MedSem-DeID · MeMa validation | disease classification accuracy: 69.94% / 71.35% -> 86.70% | Helped |
| Use diffusion instead of ViT (MAE) as the medical semantic encoder (ablation) | MedSem-DeID · MeMa validation | disease classification accuracy: 85.26% -> 86.70% | Helped |
| Increase the medical-loss weight (ablation) | MedSem-DeID · MeMa validation | disease classification accuracy | Mixed |
Surprising findings
- Identity-based face recognition ID-Dis is highest for RiDDLE, which sacrifices medical details, showing privacy and medical utility trade-off.
- Injecting public face attributes via IP-Adapter reduced identity leakage from 71.83% to 1.27% on average.
Limitations
- Dataset focuses only on eye disease manifestations.
- Patient data is from a single hospital (SNPH) for generation; real-world evaluation on a single other hospital with 129 patients.
- Some medical-utility metrics trade off against identity protection (e.g., reduced ID-Dis when medical loss weight increases).
Experimental setup
Datasets: MeMa training set Real patient photos (SNPH) FFHQ CASIA VGGFace2 LFW Real-ECXHCSU
Benchmarks: MeMa validation set MeMa-Seg Real-ECXHCSU LFW
Metrics: Top-1 classification accuracy Dice Jaccard Cohen's Kappa ID-Dis face matching rate landmark error gaze error PSNR LPIPS
Baselines: DeepPrivacy CIAGAN Disguise Password Personal RiDDLE Digital mask
Builds on: Stable Diffusion IP-Adapter LoRA RiDDLE Password
Closest papers by meaning
- CARES: A Comprehensive Benchmark of Trustworthiness in Medical Vision Language Models (NeurIPS 2024 · similarity 0.83)
- Towards All-in-One Medical Image Re-Identification (CVPR 2025 · similarity 0.83)
- How Private is Private? A Comparative Study for Face De-Identification (NeurIPS 2026 · similarity 0.82)
- DrVD-Bench: Do Vision-Language Models Reason Like Human Doctors in Medical Image Diagnosis? (NeurIPS 2025 · similarity 0.82)
- Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline (CVPR 2025 · similarity 0.82)
- Medical thinking with multiple images (ICLR 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.