期刊文献+

基于Dense Connected深度卷积神经网络的自动视网膜血管分割方法 被引量:3

Automatic Retinal Vascular Segmentation Method based on Densely Connected Convolution Neural Network
下载PDF
导出
摘要 深度卷积神经网络(DCNN)在自然图像分类和分割问题中具有优越的性能。眼底视网膜血管作为可无创直接观察到的血管,对其结构的分析是眼科病变诊断的重要依据之一。如毛细血管增生等变化为糖尿病等眼科疾病的诊断提供了重要的指导意义。因此,如何正确高效地分割视网膜血管成为一种临床需求。在不使用任何前后期处理的条件下,提出一种基于Densely Connect深度卷积神经网络的自动视网膜血管分割方法。方法通过使用稠密连接(Densely connect),批规范化(Batch Normalization)等技术构建一种新型的深度卷积神经网络,并结合带孔卷积(Dilated convolution)增加网络分割精度,在更少人为处理的情况下提高视网膜血管的分割性能。在对比实验中,提出网络的平均精确度,敏感度,特异性达到0.9617,0.7325,0.9839,像素级的AUC指标达到0.978,优于对比的机器学习方法和对比深度卷积神经网络。验证了所提方法在视网膜血管分割中的有效性。 Deep convolutional neural network (DCNN)has shown its superior performance in image classification and seg-mentation problems. It has been extensively studied and also promotes the medical image segmentation development. Fun-dus retinal blood vessels are non-invasive directly observed blood vessel, which can provide one of the most important ev-idences for the diagnosis of ophthalmic diseases. For example, capillary proliferation provides important guidance for the diagnosis of ocular diseases such as diabetes. Therefore, a correct and efficient method of retinal blood vessel segmenta- tion becomes a clinical requirement. In this paper, we propose a method of densely connected-based convolution neural network for retinal blood vessel segmentation. The proposed innovative network employs densely connect for reusing fea- tures and enhancing feature delivery. This method uses batch normalization enabling the network to converge to better re- suits with less time. Combined with the dilated convolution, the network can get more accurate segmentation results. By this method, we can gain a better performance without using pre-processing and post-processing. Through the comparison experiments with traditional machine learning methods and other DCNN segmentation methods, the proposed method can achieve better average accuracy and sensitivity. The specificity reached 0.9617,0. 7325,0.9839, and a pixel wide AUC reached 0. 978. Experimental results demonstrate the effectiveness and efficiency of our method.
作者 唐明轩 李孝杰 周激流 TANG Ming-xuan , LI Xiao-jie, ZHOU Ji-liu(College of Computer Science and Technology, Chengdu University of Information Technology, Chengdu 610225, china)
出处 《成都信息工程大学学报》 2018年第5期525-530,共6页 Journal of Chengdu University of Information Technology
基金 成都信息工程大学科研基金(KYTZ201608)
关键词 深度学习 卷积神经网络 图像分割 视网膜血管 deep learning convolutional neural network image segmentation retinal blood vessels
  • 相关文献

参考文献1

二级参考文献1

共引文献24

同被引文献10

引证文献3

二级引证文献2

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部