| 邓小波,刘奇,陈曦,何柯辰,全美霖,刘艳丽.基于通道注意力双路径架构网络分割视网膜血管[J].中国医学影像技术,2021,37(10):1543~1547 |
| 基于通道注意力双路径架构网络分割视网膜血管 |
| Retinal vessel segmentation based on dual-path channel attention network |
| 投稿时间：2021-03-17 修订日期：2021-08-18 |
| DOI：10.13929/j.issn.1003-3289.2021.10.026 |
| 中文关键词: 视网膜血管 双路径 注意力机制 深度学习 |
| 英文关键词:retinal vessels dual path attention mechanism deep learning |
| 基金项目: |
|
| 摘要点击次数: 1886 |
| 全文下载次数: 1029 |
| 中文摘要: |
| 目的 评价基于通道注意力的双路径架构网络（DPCA-Net）算法分割视网膜血管的效果。方法 基于DRIVE及CHASE_DB1公开数据集，通过在网络中引入通道注意力机制，融合于主路径与次路径网络中提取的特征，构建端-端DPCA-Net视网膜血管分割深度学习体系结构，并构建未引入通道注意力机制的双路径架构网络（DP-Net）算法，评价其分割视网膜血管的效果。结果 DPCA-Net算法可正确识别中央血管反射区中的血管；亮斑区大部分血管被正确识别，背景干扰区中仅小部分背景被认为是血管，黑斑区部分形状类似血管的黑斑被认为是血管。DPCA-Net算法分割DRIVE/CHASE_DB1数据集中视网膜血管的准确率为95.58%/96.34%，敏感度为80.37%/77.70%，特异度为97.80%/98.22%，F1值为82.24%/79.55%；除基于DRIVE数据集的敏感度之外均高于DP-Net算法。结论 相比DP-Net算法，DPCA-Net算法能学习更多血管分割特征，且对病变区域不敏感，分割视网膜血管效果较好。 |
| 英文摘要: |
| Objective To observe the effectiveness of retinal vessel segmentation based on dual path-based channel attention network (DPCA-Net) algorithm. Methods Based on two public fundus databases, DRIVE and CHASE_DB1, an end-to-end DPCA-Net deep learning architecture was constructed to segment the retina vessel through introducing the channel attention mechanism into the network and integrating the features extracted from the primary path and secondary path network. In addition, a dual path architecture network (DP-Net) algorithm was constructed without channel attention mechanism. The effectiveness of DPCA-Net and DP-Net algorithm for retinal vessel segmentation were evaluated. Results DPCA-Net could correctly identify vessels in the central vessel reflex area. In the bright spot area, most of the blood vessels were correctly identified, while in the background interference area, only a small part of the background was considered as blood vessels, whereas in the black spot area, some black spots with a shape similar to blood vessels were considered blood vessels by DPCA-Net. The accuracy, sensitivity, specificity and F1 values of DPCA-Net algorithm for retinal vessel segmentation in the DRIVE/CHASE_DB1 dataset was 95.58%/96.34%, 80.37%/77.70%, 97.80%/98.22% and 82.24%/79.55%, higher than those of DP-Net algorithm except of the sensitivity based on DRIVE dataset. Conclusion Compared with DP-Net, DPCA-Net could learn more vessel segmentation features, which was not sensitive to the lesion area and had good effectiveness for retinal vessel segmentation. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
