| 刘少鹏,洪佳明,梁杰鹏,贾西平,欧阳佳,印鉴.面向医学图像分割的半监督条件生成对抗网络.软件学报,2020,31(8):2588-2602 |
| 面向医学图像分割的半监督条件生成对抗网络 |
| Medical Image Segmentation Using Semi-supervised Conditional Generative Adversarial Nets |
| 投稿时间：2018-11-14 修订日期：2019-01-15 |
| DOI：10.13328/j.cnki.jos.005860 |
| 中文关键词: 医学图像 深度学习 生成对抗网络 半监督学习 青光眼筛查 |
| 英文关键词:medical image deep learning generative adversarial nets semi-supervised learning glaucoma screening |
| 基金项目:国家自然科学基金（61472453，61702119，U1401256，U1501252，U1611264，U1711261，U1711262）；广东省自然科学基金（2019A1515012048，2015A030310312，2014A030309013）；广东省教育厅青年创新人才项目（2017KQNCX117，2015KQNCX084）；广州市科技计划（201802010029）；广东省大数据分析与处理重点实验室开放基金（201802） |
|
| 摘要点击次数: 526 |
| 全文下载次数: 859 |
| 中文摘要: |
| 医学图像分割是计算机辅助诊断的关键技术.青光眼作为全球第二大致盲眼病，其早期筛查和临床诊断依赖于眼底图的视盘和视杯的准确分割.但传统的视盘和视杯分割方法采用人工构建特征，模型泛化能力差.近年来，基于卷积神经网络的端对端学习模型可通过自动发现特征来分割视盘和视杯，但由于标注样本有限，模型难以训练.提出一个基于半监督条件生成对抗网络的视盘和视杯两阶段分割模型——CDR-GANs.该模型的每个分割阶段均由语义分割网络、生成器和判别器构成，通过对抗学习，判别器引导语义分割网络和生成器学习眼底图及其分割图的联合概率分布.在真实数据集ORIGA上的实验结果表明，CDR-GANs在均交并比（mean intersection over union，简称MIoU）、CDR绝对误差（absolute CDR error）和实际分割效果这些指标上明显优于现有模型. |
| 英文摘要: |
| Medical image segmentation is a key technology in computer aided diagnosis. As a widespread eye disease, glaucoma may cause permanent loss in vision and its screening and diagnosis requires accurate segmentation of optic cup and disc from fundus images. Most traditional computer vision methods segment optic cup and disc with artificial features lead to limited generalization ability. While the end-to-end learning models based on convolutional neural networks focus on optic disc and cup segmentation using automatically detected features, but fail to tackle the lack of labeled samples, thus the segmentation performance is still barely satisfactory. This study proposes an effective two-stage optic disc and cup segmentation method based on semi-supervised conditional generative adversarial nets, namely CDR- GANs. Each stage builds upon three players—A segmentation net, a generator, and a discriminator, where the segmentation net and generator concentrate on learning the conditional distributions between fundus images and their corresponding segmentation maps, and the discriminator distinguishes whether the image-label pairs come from the empirical joint distribution. The extensive experiments show that the proposed method achieves state-of-the-art optic cup and disc segmentation results on ORIGA dataset. |
|
HTML 下载PDF全文 查看/发表评论 下载PDF阅读器 |
