Synthetic artificial intelligence using generative adversarial network for retinal imaging in detection of age-related macular degeneration

Zhaoran Wang, Gilbert Lim, Wei Yan Ng, Tien‐En Tan, Jane Lim, Sing Hui Lim, Valencia Hui Xian Foo, Joshua Lim, Laura Gutiérrez, Feihui Zheng, Nan Liu, Gavin Siew Wei Tan, Ching‐Yu Cheng, Chui Ming Gemmy Cheung, Tien Yin Wong, Daniel Shu Wei TingFrontiers in Medicine 2023 · DOI

Abstract

The introduction of HITL training increased the percentage of synthetic images with AMD lesions, despite the limited number of AMD images in the initial training dataset. Qualitatively, the synthesized images have been proven to be robust in that our residents had limited ability to distinguish real from synthetic ones, as evidenced by an overall accuracy of 0.66 (95% CI: 0.61-0.66) and Cohen's kappa of 0.320. For the non-referable AMD classes (no or early AMD), the accuracy was only 0.51. With the objective scale, the overall accuracy improved to 0.72. In conclusion, GAN models built with HITL training are capable of producing realistic-looking fundus images that could fool human experts, while our objective realness scale based on broken vessels can help identifying the synthetic fundus photos.

Introduction

Deep learning (DL) has revolutionized medical imaging, but its "hunger" for massive, high-quality datasets remains a major bottleneck. In the context of age-related macular degeneration (AMD)—a leading cause of global vision impairment—advanced disease phenotypes are relatively rare in general screening populations. Furthermore, strict patient privacy regulations (like HIPAA) make sharing large-scale medical datasets difficult.

Generative Adversarial Networks (GANs) offer a potential solution by creating synthetic data to augment small datasets. However, standard GANs often struggle to produce the specific, complex pathological features (like drusen or geographic atrophy) required for clinical validity. This paper tackles the challenge of generating realistic, diverse AMD images and establishing a reliable way to measure their "realness."

Methodology

The authors employed StyleGAN2 as their base architecture but introduced a Human-in-the-Loop (HITL) training strategy to compensate for data scarcity.

Preprocessing and Segmentation

To ensure the GAN focused on clinically relevant areas, U-Net models were trained to segment the macula and optic disc. The final step was to extract the circular two optic disc diameters (DD) macula region, as defined by the macula center and the radius of 2DD.

Iterative HITL Training

Because advanced AMD images were scarce in the initial dataset, the researchers manually selected the best synthetic images from early training rounds (AMDGAN v1.0 and v2.0) and fed them back into the training loop. This "guided" the model toward producing more accurate disease features.

Overview of the preprocessing segmentation and the iterative Human-in-the-Loop (HITL) training workflow for AMDGAN development.

Datasets and Evaluation

The study utilized the Singapore Integrated Diabetic Retinopathy Program (SiDRP) dataset, which was highly imbalanced, with only 151 "advanced AMD" images compared to over 90,000 "no AMD" images.

Evaluation metrics included:

  • FID & SSIM: Standard computer vision metrics to ensure image quality and diversity.
  • Subjective Grading: Four ophthalmology residents attempted to distinguish 150 real images from 150 synthetic ones.
  • Objective Realness Scale: The authors proposed a new scale (0–2) based on the frequency of "broken vessels"—a common artifact in synthetic retinal images where vascular structures abruptly terminate.

Key Results and Insights

  • HITL as a "Class Balancer": Human intervention compensated for extreme data scarcity. By manually curating the GAN's output, the authors increased the number of synthetic images with advanced AMD features from zero to 543.
  • The Turing Test for Retinas: For early-stage AMD, the synthetic images were virtually indistinguishable from real ones (accuracy was 0.51, essentially a coin flip).
  • Objective Improvement: When residents used the "broken vessel" scale, their accuracy in identifying fakes improved from 0.66 to 0.72, and the Cohen’s kappa rose from 0.320 to 0.430.
  • Diversity: The AMDGAN v3.0 achieved a Frechet Inception Distance (FID) of 6.8084, indicating high image quality and a distribution close to the real dataset.

Conclusion

GAN models built with HITL training are capable of producing realistic-looking fundus images that can fool human experts. While the HITL process is labor-intensive, it provides a valuable "sandbox" for researchers to test models without touching sensitive patient data. Future work may involve more advanced architectures like StyleGAN3 or Diffusion Models to further improve vascular connectivity.

Summary updated

Abstract

The introduction of HITL training increased the percentage of synthetic images with AMD lesions, despite the limited number of AMD images in the initial training dataset. Qualitatively, the synthesized images have been proven to be robust in that our residents had limited ability to distinguish real from synthetic ones, as evidenced by an overall accuracy of 0.66 (95% CI: 0.61-0.66) and Cohen's kappa of 0.320. For the non-referable AMD classes (no or early AMD), the accuracy was only 0.51. With the objective scale, the overall accuracy improved to 0.72. In conclusion, GAN models built with HITL training are capable of producing realistic-looking fundus images that could fool human experts, while our objective realness scale based on broken vessels can help identifying the synthetic fundus photos.
BibTeX
@inproceedings{wang2023synthetic,
  author = {Zhaoran Wang and Gilbert Lim and Wei Yan Ng and Tien‐En Tan and Jane Lim and Sing Hui Lim and Valencia Hui Xian Foo and Joshua Lim and Laura Gutiérrez and Feihui Zheng and Nan Liu and Gavin Siew Wei Tan and Ching‐Yu Cheng and Chui Ming Gemmy Cheung and Tien Yin Wong and Daniel Shu Wei Ting},
  title = {{Synthetic artificial intelligence using generative adversarial network for retinal imaging in detection of age-related macular degeneration}},
  booktitle = {Frontiers in Medicine},
  year = {2023},
  doi = {10.3389/fmed.2023.1184892},
}
Figure