SGW-GAN: Sliced Gromov-Wasserstein Guided GANs for Retinal Fundus Image Enhancement

TL;DR

This work tackles unpaired retinal fundus image enhancement while preserving clinically meaningful intra-class geometry. It introduces SGW-GAN, which embeds the efficient Sliced Gromov-Wasserstein (SGW) discrepancy into a GAN framework, using a frozen retinal encoder to compute relational regularization between enhanced and high-quality distributions. Empirical results on EyeQ show SGW-GAN achieves superior diabetic retinopathy grading and the lowest GW discrepancy, indicating preserved disease-relevant structure despite improved visual quality. The approach offers a practical, clinically faithful alternative to traditional perceptual-only enhancement with potential applicability to other medical imaging domains.

Abstract

Retinal fundus photography is indispensable for ophthalmic screening and diagnosis, yet image quality is often degraded by noise, artifacts, and uneven illumination. Recent GAN- and diffusion-based enhancement methods improve perceptual quality by aligning degraded images with high-quality distributions, but our analysis shows that this focus can distort intra-class geometry: clinically related samples become dispersed, disease-class boundaries blur, and downstream tasks such as grading or lesion detection are harmed. The Gromov Wasserstein (GW) discrepancy offers a principled solution by aligning distributions through internal pairwise distances, naturally preserving intra-class structure, but its high computational cost restricts practical use. To overcome this, we propose SGW-GAN, the first framework to incorporate Sliced GW (SGW) into retinal image enhancement. SGW approximates GW via random projections, retaining relational fidelity while greatly reducing cost. Experiments on public datasets show that SGW-GAN produces visually compelling enhancements, achieves superior diabetic retinopathy grading, and reports the lowest GW discrepancy across disease labels, demonstrating both efficiency and clinical fidelity for unpaired medical image enhancement.

Figures (5)

• Figure 1: t-SNE visualizations of feature distributions across 2 disease labels, calculated from encoded feature using RetFound zhou2023foundation: (a) Low-quality images already show reasonable separation between labels; (b) After enhancement by OTEGAN zhu2023otre, the class separation become blurred, distorting the intra-space geometry.
• Figure 2: Illustration of how SGW accelerates the GW computing process by projecting features onto multiple random one-dimensional directions, computing one-dimensional GW in each subspace, and aggregating results with multiple projections. This reduces computational cost by orders of magnitude while preserving relational fidelity.
• Figure 3: Overview of the SGW-GAN framework. The generator produces enhanced images from low-quality inputs, while the discriminator distinguishes enhanced outputs from real high-quality images under a WGAN objective. The generator is optimized with three complementary losses: (1) RMSE regularization to preserve the eyeball structure and prevent over-enhancement, (2) SGW discrepancy to maintain intra-class relational geometry, and (3) adversarial loss to improve realism. Training alternates between updating the discriminator and the generator, ensuring that enhanced images achieve both visual quality and structural alignment with the high-quality distribution.
• Figure 4: Proposed generator structure: a U-Net with residual connection and channel attention adapted from zhu2023optimalwang2022optimal.
• Figure 5: Representative examples of retinal image enhancement. Each triplet shows the low-quality input (top row), the ground-truth high-quality image (middle row), and the enhanced output by our method (bottom row). Despite slightly lower SSIM/PSNR values compared to previous approaches, our method preserves fine vessels and overall retinal structure, while effectively removing black dots and artifacts, leading to clearer and more clinically usable images.