DeepSynthRetina: A Vision-Based Model for Synthetic Fundus Image Generation

The identification of ocular diseases such as Age-related Diabetic Retinopathy (DR), Macular Degeneration (AMD), and Glaucoma heavily relies on retinal imaging techniques. Recent progress in medical imaging has incorporated automatic diagnosis algorithms, primarily leveraging deep learning technologies. However, the effectiveness of these methods hinges on large datasets for model training, presenting challenges due to the scarcity, limited availability, and high acquisition costs associated with medical data. Consequently, there is an urgent demand for data generation, particularly within the domain of healthcare imaging. This research addresses issues related to the shortage and availability of medical data, as well as the inherent challenge of data imbalance that impedes the development of high-performance models. The suggested method leverages a pre-trained Texture Model to efficiently capture and analyze structural details from retinal images. A Generative Adversarial Network (GAN) is then used to train the model, allowing for the creation of images based on the extracted texture model. This methodology serves as a practical solution to overcome data scarcity and imbalance, contributing to the generation of synthetic fundus images with the potential to enhance the robustness and generalization capabilities of automated diagnosis algorithms in ophthalmic medical imaging. Extensive experimental results demonstrate that our DeepSynthRetina method provides valuable insights into the practical utility of synthesized images in classification scenarios, significantly outperforming the Baseline model for DR grading in public datasets.