Privacy-preserving collaborative transfer learning with generative adversarial networks (CTL-GAN) for classification of dry eye disease
Abstract
This study investigated whether a collaborative transfer learning generative adversarial network (CTL-GAN) can generate synthetic fluorescein corneal images and enhance multiclass dry eye disease (DED) grading across institutions. We retrospectively included 1,600 fluorescein-stained anterior segment images from patients with DED at Samsung Medical Center (SMC) and 260 images from the public Sun Yat-sen University (SYSU) dataset, graded on the Oxford scale. A StyleGAN3 generator trained on SMC data produced synthetic images. Pretrained generator weights were transferred to a simulated site represented by the SYSU dataset and fine-tuned on local images, without exchanging raw images. DenseNet classifiers trained with GAN-based augmentation were compared with a benchmark model trained on real images only. GAN-based augmentation improved validation accuracy on SMC data from 64.2% to 75.5%, and transferring pretrained weights increased SYSU validation accuracy from 41.4% to 66.3%. Peak signal-to-noise ratio (PSNR 13.95 ± 1.54 dB), structural similarity index (SSIM 0.428 ± 0.111), and learned perceptual image patch similarity (LPIPS 0.510 ± 0.047) showed low similarity to real images. In a blinded 10-item multiple-choice realism survey, 37 ophthalmologists answered correctly in 28.4% of cases, against a 20% chance level. CTL-GAN indicates that image synthesis from shared generator weights can improve external DED grading without exchanging patient images.
Related
- AI expert collaborative generative model for predicting pathologic response to neoadjuvant chemoimmunotherapy in NSCLC
- Automated diagnosis of keratitis from low-quality slit-lamp images using an improved generative adversarial network
- Human-AI collaborative generative AI for comfortable care: a doctor–patient–AI feedback loop for pre-operative distress and workforce shortages
- Unmasking adversarial attacks using a robust XAI-driven approach for secure medical image classification
- Prediction of lncRNA–cervical cancer associations using a multi-view collaborative variational autoencoder with knowledge distillation transfer learning
BibTeX
@article{kim2026privacypreserving,
title = {{Privacy-preserving collaborative transfer learning with generative adversarial networks (CTL-GAN) for classification of dry eye disease}},
author = {Bongsang Kim and Yeo Kyoung Won and Hokyun Kim and Jiho Cha and Dong Hui Lim and Jooyoung Jeon},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-76373-4},
url = {https://www.nature.com/articles/s41598-026-76373-4},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Privacy-preserving collaborative transfer learning with generative adversarial networks (CTL-GAN) for classification of dry eye disease
AU - Bongsang Kim
AU - Yeo Kyoung Won
AU - Hokyun Kim
AU - Jiho Cha
AU - Dong Hui Lim
AU - Jooyoung Jeon
PY - 2026
DA - 2026/10/20
JO - Scientific Reports
DO - 10.1038/s41598-026-76373-4
UR - https://www.nature.com/articles/s41598-026-76373-4
AB - This study investigated whether a collaborative transfer learning generative adversarial network (CTL-GAN) can generate synthetic fluorescein corneal images and enhance multiclass dry eye disease (DED) grading across institutions. We retrospectively included 1,600 fluorescein-stained anterior segment images from patients with DED at Samsung Medical Center (SMC) and 260 images from the public Sun Yat-sen University (SYSU) dataset, graded on the Oxford scale. A StyleGAN3 generator trained on SMC data produced synthetic images. Pretrained generator weights were transferred to a simulated site represented by the SYSU dataset and fine-tuned on local images, without exchanging raw images. DenseNet classifiers trained with GAN-based augmentation were compared with a benchmark model trained on real images only. GAN-based augmentation improved validation accuracy on SMC data from 64.2% to 75.5%, and transferring pretrained weights increased SYSU validation accuracy from 41.4% to 66.3%. Peak signal-to-noise ratio (PSNR 13.95 ± 1.54 dB), structural similarity index (SSIM 0.428 ± 0.111), and learned perceptual image patch similarity (LPIPS 0.510 ± 0.047) showed low similarity to real images. In a blinded 10-item multiple-choice realism survey, 37 ophthalmologists answered correctly in 28.4% of cases, against a 20% chance level. CTL-GAN indicates that image synthesis from shared generator weights can improve external DED grading without exchanging patient images.
SN - 2045-2322
ER -