Early experience of adopting a generative diffusion model for the synthesis of fundus photographs

This study was based on a publicly accessible and deidentified FP image database, which was released by a previous study. To reduce the noise of different borders, we selected FPs where all boundaries of the circular mask were intact. Finally, 1000 FPs were used to train the diffusion model. Therefore, this study was exempt from ethical review according to the guidelines by Korea National Institute for Bioethics Policy. All the procedures were performed in accordance with the ethical standards of the 1964 Declaration of Helsinki and its later amendments. As shown in Figure 1, we trained the DDPM based on U-Net backbone architecture, which is the most popular form of generative diffusion model. After training, serial multiple denoising U-Nets can generate FPs using random noise seeds. The input image size was set to a pixel resolution of 128 × 128. This was the maximum resolution required to finalize the proper training with our computational resources. We set the time step to the default value of 1000 because fewer numbers produce severely noisy or blurred synthetic images. All codes for the implementation of the DDPM are available on the webpage (https://github.com/lucidrains/denoising-diffusion-pytorch). For reproducibility, our FP images and modified codes, which can be implemented in Google Colaboratory, were released in the data repository.本研究依托于一项既往研究发布的可公开获取且经去标识化处理的FP图像数据库。为消除不同边界带来的噪声干扰，我们筛选出所有圆形掩码（circular mask）边界均完整的FP图像，最终选取1000幅FP图像用于训练扩散模型。因此，根据韩国国家生物伦理政策研究院（Korea National Institute for Bioethics Policy）的指南，本研究可豁免伦理审查。本研究所有操作均符合1964年《赫尔辛基宣言》及其后续修订版的伦理标准。如图1所示，我们基于当前最主流的生成式扩散模型架构——U-Net主干网络，训练了去噪扩散概率模型（DDPM，Denoising Diffusion Probabilistic Model）。训练完成后，串行多阶段去噪U-Net可通过随机噪声种子生成FP图像。输入图像的像素分辨率设为128×128，这是在当前计算资源下可完成有效训练的最大分辨率。我们将时间步长设为默认的1000，若步长更小则会生成噪声过重或模糊的合成图像。本研究中实现DDPM的全部代码已公开至网页（https://github.com/lucidrains/denoising-diffusion-pytorch）。为确保研究可复现，我们将本次研究使用的FP图像及可在谷歌协作平台（Google Colaboratory）上运行的修改版代码上传至了数据仓库。