Self-supervised Domain Adaptation for Breaking the Limits of Low-quality Fundus Image Quality Enhancement

TL;DR

This work tackles the problem of enhancing low-quality retinal fundus images without relying on high-quality references or paired data. It introduces DASQE, a fully self-supervised domain-adaptation framework that constructs patch-wise domains, and disentangles content, low-quality factors, and target style to produce quality-consistent fundus images. By leveraging self-supervision and adversarial losses in both latent and image spaces, it achieves state-of-the-art PSNR/SSIM on EyeQ and Messidor without high-quality guidance, and improves downstream tasks such as vessel/lesion segmentation and DR grading. The approach offers practical clinical value by reducing reliance on high-quality data and demonstrating improved robustness across datasets and tasks, with potential extension to other medical imaging domains.

Abstract

Retinal fundus images have been applied for the diagnosis and screening of eye diseases, such as Diabetic Retinopathy (DR) or Diabetic Macular Edema (DME). However, both low-quality fundus images and style inconsistency potentially increase uncertainty in the diagnosis of fundus disease and even lead to misdiagnosis by ophthalmologists. Most of the existing image enhancement methods mainly focus on improving the image quality by leveraging the guidance of high-quality images, which is difficult to be collected in medical applications. In this paper, we tackle image quality enhancement in a fully unsupervised setting, i.e., neither paired images nor high-quality images. To this end, we explore the potential of the self-supervised task for improving the quality of fundus images without the requirement of high-quality reference images. Specifically, we construct multiple patch-wise domains via an auxiliary pre-trained quality assessment network and a style clustering. To achieve robust low-quality image enhancement and address style inconsistency, we formulate two self-supervised domain adaptation tasks to disentangle the features of image content, low-quality factor and style information by exploring intrinsic supervision signals within the low-quality images. Extensive experiments are conducted on EyeQ and Messidor datasets, and results show that our DASQE method achieves new state-of-the-art performance when only low-quality images are available.

Figures (7)

• Figure 1: Some cases of low-quality fundus images, where the red box and the yellow box mark the different quality regions in the low-quality images, respectively. The main low-quality factors include uneven illumination, noticeable blur and artifacts.
• Figure 2: Comparison of different image quality enhancement schemes. Previous fundus image quality enhancement methods (A) require paired images of high-quality and low-quality as training data, which are often difficult to acquire. Although these semi-supervised and unsupervised methods (B) and (C) eliminate the need for paired images, they still need the guidance of high-quality images. (D) In contrast, our method is a truly unsupervised image quality enhancement, requiring neither paired fundus images nor high-quality images.
• Figure 3: The overall architecture of the proposed framework. It involves two stages: Construction of patch-wise image domains:The low-quality input images $\boldsymbol{X}$ are first serialized as the patch-wise fundus images $\boldsymbol{X}=\left\{x^1, \ldots, x^m\right\}$. With the help of the pre-trained quality assessment network, the patch-wise fundus images are divided into low-quality and high-quality domains, e.g. $\boldsymbol{X_L},\boldsymbol{X_H}$. A style clustering is performed on $\boldsymbol{X_H}$, and a target style domain $\boldsymbol{X^T_H}$ is determined. Decoupling of multiple features: We factorize content-related, quality-related and style-related for patch-wise domains through the supervisions derived from the image data itself to enhance the image quality and align style features. The framework involves content-related encoders {$\boldsymbol{E}^C_{x_L}$, $\boldsymbol{E}^C_{x^S_H}$,$\boldsymbol{E}^C_{x^T_H}$} for extracting latent content embedding, quality-related encoder $\boldsymbol{E}^R_{x_L}$ for extracting low-quality factor embedding, target style-related encoder $\boldsymbol{E}^S_{x^T_H}$ for extracting target style embedding, low-/high-quality image generators $\boldsymbol{G}_L$ and $\boldsymbol{G}^S_H$ for generating different quality images, and high-quality target style image generator high-quality style images.
• Figure 4: Visual comparisons on the low-quality image enhancement between the proposed and other deep learning methods.
• Figure 5: Comparison of low-quality fundus image enhancement results based on the different patch sizes (a-b) and different number K of style cluster centers (c-d).