基于LDU的视网膜OCT图像分层分割研究

Layer segmentation of retinal OCT images based on layer dense block U-Net (LDU)

U-Net是近期流行的视网膜OCT图像分层分割方法,但是在具有复杂纹理信息的图像区域上会出现较大分割误差。提出一种带有层密集块的U型神经网络(LDU)。通过引入图像特征信息重用的层密集块,提取更有效的图像多尺度特征,从而提高视网膜OCT图像的分层分割精度。利用杜克大学公开的两层视网膜OCT数据库对LDU与传统U-Net方法进行对比分析。测试验证结果表明该方法可提升人眼特别是AMD病眼视网膜OCT图像分层分割的性能,提高视网膜各层定量化分析的准确性。

U-Net has been a popular method nowadays in the field of layer segmentation of retinal OCT images, but large segmentation errors still exist in image regions with complex texture information. In this paper, we propose a layer dense block-based U-type neural network(LDU) for this purpose. By introducing layer dense blocks for image feature information reuse, more effective multi-scale features of images are extracted, thus improving the accuracy of layered segmentation. The LDU is compared to the conventional U-Net method using the publicly available two-layer retinal OCT database from the Duke University. The test results show that this method could improve the performance of layered segmentation of retinal OCT images in human eyes, especially those with AMD, which consequently improved the accuracy of quantitative analysis for each retinal layer.