Lesion-Conditioned Fundus Image Synthesis for Diabetic Retinopathy Using Pix2Pix Conditional GANs

Abstract

Deep learning models for diabetic retinopathy (DR) demonstrate superior performance when trained on an extensive, classbalanced fundus dataset. However, widely used collections, such as EyePACS, exhibit significant class imbalances, particularly with the underrepresentation of severe stages. We propose a lesion-conditioned, pathology-aware synthesis framework to augment DR fundus images using a Pix2Pix-based conditional generative adversarial network. Lesion structures are incorporated through segmentation masks for four clinically significant lesion types—microaneurysms, hemorrhages, hard exudates, and soft exudates—derived from pixellevel annotations and/or a trained U-Net segmenter, which serves as conditioning inputs for the generator. To enhance data availability for minority grades, we implemented a minimum-real subsampling baseline and augmented each grade with lesion-fidelity-filtered synthetic images using a mask cycling strategy. We present qualitative results that demonstrate spatial consistency between the conditioned lesion masks and the synthesized appearance, and we include quantitative image-quality descriptors (Fréchet Inception Distance), discussing their limitations in small, low-diversity medical datasets. Finally, we evaluated the downstream utility through a real-only versus real + synthetic robustness study across three random seeds, reporting mean ± std and explicitly noting the small held-out test subset used in this paper. Overall, the proposed framework facilitates controlled, lesion-preserving synthesis for class balancing and offers an auditable augmentation protocol for DR model development.



Received: 5 December 2025 | Revised: 27 March 2026 | Accepted: 18 May 2026



Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.



Data Availability Statement

The IDRiD datasets that support the findings of this study are openly available at https://idrid.grand-challenge.org/Data/. The EyePACS datasets that support the findings of this study are openly available at https://www.kaggle.com/c/diabetic-retinopathy-detection/data.


Author Contribution Statement

Anju Asokan: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – original draft, Visualization. Manju Bhaskara Panickar Radha: Writing – review & editing, Supervision, Project administration.