基于分步生成模型的视网膜眼底图像合成

A multi-step generation method for synthesizing retinal fundus images

由于目前带标注的医学图像稀缺且不平衡,论文针对带标注的视网膜眼底图的生成提出一种有效的分步生成方法首先训练一个标注生成对抗网络用于生成血管树标注图像,然后训练一个标注转换对抗网络用于将血管树转换为视网膜眼底图。两步训练完后实现输入一段噪声即可同时生成血管树和视网膜眼底图。生成的标注和医学图像有合理的解剖结构,应用于分割任务中也表现出与真实训练集相近的精度。以上结果表明本文提出的分步生成模型生成的数据能在一定程度上缓解医学图像及其标注不足的问题。

Considering the scarcity of medical images with annotations,we propose an effective multi-step generation method for retinal fundus images with corresponding vascular tree labels in this paper.Firstly,a generation adversarial network is trained to generate vascular tree images.Then a second adversarial net-work is trained to convert the vascular tree into a retinal fundus image.With these two steps,a noise vector can be input to generate a vascular tree and a matched retinal fundus image.The generated images have reasonable anatomical structures.Similar accuracy is obtained in segmentation task when using generated images for training compared with using real dataset.These results illustrate that generating data by the proposed method can alleviate the problem of insufficient medical images and annotations to some extent.