Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust deep learning for eye fundus images: Bridging real and synthetic data for enhancing generalization

Guilherme C. Oliveira, Gustavo H. Rosa, Daniel C. G. Pedronette, João P. Papa, Himeesh Kumar, Leandro A. Passos, Dinesh Kumar

TL;DR

The paper tackles the challenge of limited, imbalanced fundus image datasets for AMD detection by proposing a pipeline that combines retinal image quality assessment with StyleGAN2-ADA to synthesize realistic images. It conducts a comprehensive comparison of GANs, demonstrates that StyleGAN2-ADA yields the best image quality (FID = 166.17) and that clinicians struggle to distinguish synthetic from real images, and shows that a ResNet-18 trained on mixed real and synthetic data achieves strong AMD detection with good generalization to the STARE external set. The approach yields AMD classification performance that rivals or exceeds human experts, and the authors provide open-source code and a web-based tool to facilitate broad adoption and further research. This work advances generalization in medical-imaging pipelines through high-quality synthetic data and accessible deployment options for AMD screening.

Abstract

Deep learning applications for assessing medical images are limited because the datasets are often small and imbalanced. The use of synthetic data has been proposed in the literature, but neither a robust comparison of the different methods nor generalizability has been reported. Our approach integrates a retinal image quality assessment model and StyleGAN2 architecture to enhance Age-related Macular Degeneration (AMD) detection capabilities and improve generalizability. This work compares ten different Generative Adversarial Network (GAN) architectures to generate synthetic eye-fundus images with and without AMD. We combined subsets of three public databases (iChallenge-AMD, ODIR-2019, and RIADD) to form a single training and test set. We employed the STARE dataset for external validation, ensuring a comprehensive assessment of the proposed approach. The results show that StyleGAN2 reached the lowest Frechet Inception Distance (166.17), and clinicians could not accurately differentiate between real and synthetic images. ResNet-18 architecture obtained the best performance with 85% accuracy and outperformed the two human experts (80% and 75%) in detecting AMD fundus images. The accuracy rates were 82.8% for the test set and 81.3% for the STARE dataset, demonstrating the model's generalizability. The proposed methodology for synthetic medical image generation has been validated for robustness and accuracy, with free access to its code for further research and development in this field.

Robust deep learning for eye fundus images: Bridging real and synthetic data for enhancing generalization

TL;DR

The paper tackles the challenge of limited, imbalanced fundus image datasets for AMD detection by proposing a pipeline that combines retinal image quality assessment with StyleGAN2-ADA to synthesize realistic images. It conducts a comprehensive comparison of GANs, demonstrates that StyleGAN2-ADA yields the best image quality (FID = 166.17) and that clinicians struggle to distinguish synthetic from real images, and shows that a ResNet-18 trained on mixed real and synthetic data achieves strong AMD detection with good generalization to the STARE external set. The approach yields AMD classification performance that rivals or exceeds human experts, and the authors provide open-source code and a web-based tool to facilitate broad adoption and further research. This work advances generalization in medical-imaging pipelines through high-quality synthetic data and accessible deployment options for AMD screening.

Abstract

Deep learning applications for assessing medical images are limited because the datasets are often small and imbalanced. The use of synthetic data has been proposed in the literature, but neither a robust comparison of the different methods nor generalizability has been reported. Our approach integrates a retinal image quality assessment model and StyleGAN2 architecture to enhance Age-related Macular Degeneration (AMD) detection capabilities and improve generalizability. This work compares ten different Generative Adversarial Network (GAN) architectures to generate synthetic eye-fundus images with and without AMD. We combined subsets of three public databases (iChallenge-AMD, ODIR-2019, and RIADD) to form a single training and test set. We employed the STARE dataset for external validation, ensuring a comprehensive assessment of the proposed approach. The results show that StyleGAN2 reached the lowest Frechet Inception Distance (166.17), and clinicians could not accurately differentiate between real and synthetic images. ResNet-18 architecture obtained the best performance with 85% accuracy and outperformed the two human experts (80% and 75%) in detecting AMD fundus images. The accuracy rates were 82.8% for the test set and 81.3% for the STARE dataset, demonstrating the model's generalizability. The proposed methodology for synthetic medical image generation has been validated for robustness and accuracy, with free access to its code for further research and development in this field.
Paper Structure (17 sections, 7 figures, 5 tables)

This paper contains 17 sections, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset
  4. Evaluation Measures
  5. Experimental Setup
  6. Results
  7. Synthetic Image Assessment
  8. Distinguishing between Synthetic and Real Images
  9. Data Augmentation Assessment
  10. Comparison between Human Experts and Deep Models
  11. Web Application
  12. Model Generalization
  13. Discussion
  14. Conclusion and Future Works
  15. Data availability
  16. ...and 2 more sections

Figures (7)

  • Figure 1: Retina fundus image positive to age-related macular degeneration identified by the presence of drusen. The image was extracted from the iChallenge-AMD dataset fu2020adam.
  • Figure 2: Sample image extracted from ODIR-2019 dataset and its corresponding transformations: (a) original image, (b) background removal using Hough Circle Transform and resizing, and (c) central cropping.
  • Figure 3: Number of images per dataset to compose the test set: (a) images positive to AMD and (b) non-AMD images.
  • Figure 4: Examples of (a) real retina images extracted from the training dataset, and (b) synthetic images generated by StyleGAN2-ADA.
  • Figure 5: Examples of synthetic and real images for AMD and Non_AMD. (a) real, positive AMD, (b) synthetic, positive AMD, (c) real, Non-AMD and (d) synthetic, non-AMD.
  • ...and 2 more figures