张思杰,方翔,魏赋.基于GAN的少样本视网膜血管分割研究*[J].电子测量与仪器学报,2021,35(11):132-142
基于GAN的少样本视网膜血管分割研究*
Research on retinal vascular segmentation based on GAN using few samples
  
DOI:
中文关键词:  视网膜血管分割  生成对抗网络  少样本学习  类不平衡
英文关键词:retina vessel segmentation  generation of adversarial network  few sample learning  class imbalance
基金项目:国家自然科学基金(30470469)项目资助
作者单位
张思杰 重庆大学微电子与通信工程学院重庆400044 
方翔 重庆大学微电子与通信工程学院重庆400044 
魏赋 重庆大学微电子与通信工程学院重庆400044 
AuthorInstitution
Zhang Sijie College of Microelectronics and Communication Engineering, Chongqing University, Chongqing 400044, China 
Fang Xiang College of Microelectronics and Communication Engineering, Chongqing University, Chongqing 400044, China 
Wei Fu College of Microelectronics and Communication Engineering, Chongqing University, Chongqing 400044, China 
摘要点击次数: 494
全文下载次数: 0
中文摘要:
      视网膜血管分割是自动筛查糖尿病视网膜病变的重要步骤,当前大部分深度学习方法都使用大样本进行网络训练,但医学领域带标签样本难以获取,且存在健康人样本与患者样本不平衡问题。提出了一种基于生成对抗网络的少样本视网膜血管分割方法,生成器部分对图像做反色等预处理后,通过旋转增扩充了数据集,网络部分使用U Net结构,判别器部分使用卷积神经网络。在实验阶段,在DRIVE数据集和HRF数据集上进行训练测试,训练时只使用训练集的6个样本,测试时使用全部测试集样本,最终在两个数据集下的ROC曲线下面积分别达到了097和095,准确率达到了095和094。与少样本情况下的U Net相比,分割性能提升很大,表明本方法针对少样本视网膜血管分割任务确实有效。
英文摘要:
      Retinal vessel segmentation is an important step for automatic screening of diabetic retinopathy. Currently, most deep learning methods use a large number of labeled samples for network training, but it is difficult to obtain labeled samples in the medical field, and healthy samples and patient samples are imbalanced. In this paper, we have proposed a method of retinal vessel segmentation using few samples based on generating adversarial network. In the generator part, after preprocessing the image by inversing color and other methods, the dataset is expanded by rotation. The U Net structure is used in the network part and the discriminator uses CNN network. In the experimental stage, the training test was applied to DRIVE dataset and HRF dataset. Only 6 samples of the training set were used in the training step, and all the test samples were used in the test step. Finally, the area under the ROC curve of the two datasets reached 097 and 095, and the accuracy rate reached 095 and 094. Compared with U Net in condition of few samples, segmentation performance is improved greatly. It shows that this method is effective for the task of low sample segmentation in retina vessel.
查看全文  查看/发表评论  下载PDF阅读器