A CycleGAN deep learning technique for artifact reduction in fundus photography

Herein, we present a deep learning technique to remove artifacts automatically in fundus photograph. By using a CycleGAN model, we synthesize the retinal images with artifact reduction based on low-quality image, and validated this technique in the independent test dataset. This study included total 2,206 anonymized retinal images. We collected the fundus photographs without qualification, which include normal and pathologic retinal images. Images including both photograph with and without artifacts were crawled from Google image and dataset search using English keywords related to retina. The search strategy was based on the key terms “fundus photography”, “retinal image”, “artifact”, “quality assessment”, “retinal image grade”, “diabetic retinopathy”, “age-related macular degeneration”, “glaucoma”, “cataract”, and “fundus dataset”. Images with artifact were manually classified by authors. Finally, 1,146 images with artifacts and 1,060 images without artifacts were collected. The experiment process complied with the Declaration of Helsinki. This study did not require ethics committee approval; instead, researchers used open web-based and deidentified data. We used the CoLaboratory’s CycleGAN tutorial page to develop and to validate CycleGAN model, and all codes were available in the webpage (https://www.tensorflow.org/tutorials/generative/cyclegan). **This dataset may include MESSIDOR, HRF, FIRE, DRIVE, Kaggle DMR, and freely available images from Google image search. Images with and without artifacts were categorized to investigate artifact reduction.本研究提出一种可自动去除眼底照片伪影的深度学习技术。本研究基于低质量眼底图像，采用CycleGAN（CycleGAN）模型合成去除伪影后的视网膜图像，并在独立测试数据集上对该技术进行了验证。本研究共纳入2206张已匿名化的视网膜图像。本研究收集了未经筛选的眼底照片，涵盖正常视网膜图像与病理性视网膜图像。含伪影与不含伪影的图像均通过谷歌图片搜索及数据集检索获取，检索所用英文关键词与视网膜相关。本次检索采用的关键词包括：“fundus photography”、“retinal image”、“artifact”、“quality assessment”、“retinal image grade”、“diabetic retinopathy”、“age-related macular degeneration”、“glaucoma”、“cataract”及“fundus dataset”。研究人员手动对含伪影的图像进行了分类标注。最终共收集到含伪影图像1146张，不含伪影图像1060张。本研究的实验过程符合《赫尔辛基宣言》的相关规定。本研究无需伦理委员会审批，因研究人员所使用的均为公开网络获取的匿名化数据。本研究借助谷歌Colaboratory的CycleGAN教程页面开发并验证CycleGAN模型，所有代码均可在该网页（https://www.tensorflow.org/tutorials/generative/cyclegan）获取。本数据集可能包含MESSIDOR、HRF、FIRE、DRIVE、Kaggle DMR数据集及谷歌图片搜索获取的公开图像。研究对含伪影与不含伪影的图像进行分类，以探究伪影去除效果。