LLRetina:Low-Light Color Fundus Image Dataset

The LLRetina dataset comprises paired low-light (LL) and high-light (HL) color fundus images. Based on flash irradiance during imaging, the dataset is divided into two subsets: LLRetina_LL1 and LLRetina_LL2. LLRetina_LL1 includes both healthy and diseased eyes, while LLRetina_LL2 includes only healthy eyes. For LLRetina_LL1, we adopted a feature point-based image registration algorithm to estimate transformation matrices and spatially align each LL image with its corresponding HL image. The ratio between the intersection area of the unmasked regions post-registration and the unmasked area of the original low-light image is computed and pairs with a ratio below 0.9 were excluded. Additional manual inspections were performed to remove registration artifacts and regenerate the associated masks. Ultimately, a total of 630 well-aligned image pairs from 199 eyes were retained. An eye-level split is adopted to prevent data leakage. Specifically, the dataset was divided into training, validation, and test sets with 492/155, 55/17, and 83/27 image pairs / eyes, respectively. AutoMorph (Zhou et al., 2022) is used to generatevessel masks for all images, which are used for both training supervision and the proposed LFPS metric. For LLRetina_LL2, only 477 pairs of original images without image registration and quality cleaning preprocessing are provided.The dataset will be made publicly available upon acceptance of the manuscript.