曾兴晖, 许祥丛, 李晓, 王茗祎, 钟俊平, 熊红莲. 基于RAU-net的视网膜OCT图像快速自动分层研究[J]. 华南师范大学学报(自然科学版), 2021, 53(2): 1-6. doi: 10.6054/j.jscnun.2021019
引用本文: 曾兴晖, 许祥丛, 李晓, 王茗祎, 钟俊平, 熊红莲. 基于RAU-net的视网膜OCT图像快速自动分层研究[J]. 华南师范大学学报(自然科学版), 2021, 53(2): 1-6. doi: 10.6054/j.jscnun.2021019
ZENG Xinghui, XU Xiangcong, LI Xiao, WANG Mingyi, ZHONG Junping, XIONG Honglian. Quick and Automatic Segmentation of Retinal Layers in OCT Images Using Residual and Attention U-net[J]. Journal of South China Normal University (Natural Science Edition), 2021, 53(2): 1-6. doi: 10.6054/j.jscnun.2021019
Citation: ZENG Xinghui, XU Xiangcong, LI Xiao, WANG Mingyi, ZHONG Junping, XIONG Honglian. Quick and Automatic Segmentation of Retinal Layers in OCT Images Using Residual and Attention U-net[J]. Journal of South China Normal University (Natural Science Edition), 2021, 53(2): 1-6. doi: 10.6054/j.jscnun.2021019

基于RAU-net的视网膜OCT图像快速自动分层研究

Quick and Automatic Segmentation of Retinal Layers in OCT Images Using Residual and Attention U-net

  • 摘要: 利用计算机图像处理技术自动分割视网膜图像,获得各层的厚度,可对多种视网膜疾病进行直观评估. 为了快速准确地对OCT视网膜图像进行自动分层,提出一种结合RAU-net和图像搜索的视网膜图像自动分层方法. 该方法在U-net的基础上加入了残差块结构和注意力门结构,残差块结构在构建更深的网络、获取高级特征的同时,有效避免了梯度消失和梯度爆炸问题,注意力门结构突出了模型对视网膜图像重要特征的学习. 由模型预测后得到的分割结果获取9条边界的感兴趣区域,然后运用图像搜索对分层图像进行边界优化,得到精确的视网膜分层图像. 结果表明:该RAU-net算法与手动分层的误差约为1像素,且完成1帧OCT视网膜图像的分层只需要4 s. 通过RAU-net与图像搜索相结合的方法为视网膜疾病的临床诊断和治疗提供了快速准确的定量分析方法.

     

    Abstract: Computer image processing technology is used to segment the retinal image automatically and compute the thickness of each layer. The thickness can be used to assess various retinal diseases directly. In order to segment retinal layers quickly and accurately, a new segmentation method combining RAU-net (Residual and Attention U-net, RAU-net) and graph search is proposed. This method adds residual block structure and attention gate structure to the basis of U-net. The residual block structure avoids the problems of gradient disappearance and gradient explosion effectively while obtaining advanced features by building deeper networks. Models trained with attention gate structure highlights the learning of salient features of retinal images. The segmentation results obtained after the model prediction gain nine boundary regions of interest, and then the graph search method is used to optimize the layer boundary for accurate retinal layer. The results show that the error between the RAU-net algorithm and manual segmentation is just about one pixel, and it only takes 4 s to complete segmentation of an OCT retinal image. Combining RAU-net and the graph search algorithm, the method provides a fast and accurate quantitative analysis tool for the clinical diagnosis and treatment of retinal diseases.

     

/

返回文章
返回