Deep Learning-Based Style Transfer from Ultra-Widefield to Traditional Fundus Photography

This study included 451 anonymized UWF and 745 FP images. The ultra-widefield (UWF) images, which include both normal and pathologic retinal images, were based on Tsukazaki Optos Public Project. The traditional fundus photograph (FP) images were extracted from the publicly accessible database by using the Google image and Google dataset search that included English keywords related to retina. The search strategy was based on the following key terms: “fundus photography”, “retinal image”, and “fundus dataset”. The images were manually reviewed by two board-certified ophthalmologists, and blurred and low-quality images were removed to clarify the image domains. Duplicated images were also removed. Consequently, 451 images with artifacts and 745 images without artifacts were collected. The UWF images were cropped and masked after registration for CycleGAN.本研究共纳入451张匿名化超宽视野（ultra-widefield, UWF）图像与745张传统眼底照片（fundus photograph, FP）。其中涵盖正常与病理性视网膜图像的超宽视野图像源自Tsukazaki Optos公共项目；传统眼底照片则通过谷歌图片搜索与谷歌数据集搜索，以“fundus photography（眼底摄影）”“retinal image（视网膜图像）”及“fundus dataset（眼底数据集）”这三个与视网膜相关的英文关键词，从公开数据库中提取得到。所有图像均由两名经专科认证的眼科医师人工审阅，剔除模糊及低质量图像以明确图像域范围，同时移除重复图像。最终共收集到451张带有伪影的图像与745张无伪影的图像。研究人员对超宽视野图像完成配准后，对其进行裁剪与掩码处理，以用于CycleGAN模型。