光学学报, 2019, 39 (8): 0810004, 网络出版: 2019-08-07
自适应尺度信息的U型视网膜血管分割算法 下载: 1743次
U-Shaped Retinal Vessel Segmentation Algorithm Based on Adaptive Scale Information
图像处理 视网膜血管 形态学滤波 可变形卷积 空洞卷积 image processing retinal vessels morphological filtering deformable convolution dilated convolution
摘要
针对视网膜血管形态结构和尺度信息复杂多变的特点,提出一种自适应血管形态结构和尺度信息的U型视网膜血管分割算法。首先采用二维K-L(Karhunen-Loeve)变换(即霍特林变换)综合分析彩色图像三通道的频带信息,从而得到视网膜灰度图像以及多尺度形态学滤波增强血管与背景的对比度信息。然后将预处理图像经U型分割模型对图像进行端对端训练,并利用局部信息熵采样进行数据增强。该网络编码部分的密集可变形卷积结构根据上下特征层信息有效地捕捉图像中多种尺度信息和形状结构,底部金字塔型的多尺度空洞卷积扩大局部感受野,同时解码阶段带有Attention机制的反卷积网络将底层与高层特征映射有效结合,解决权重分散和图像纹理损失的问题。最后通过SoftMax激活函数得到最终的分割结果。在DRIVE(Digital Retinal Images for Vessel Extraction)与STARE(Structured Analysis of the Retina)数据集上对该算法进行了仿真,准确率分别达到97.48%与96.83%,特异性分别达到98.83%与97.75%,总体性能优于现有算法。
Abstract
In view of the complex and changeable morphological structure and scale information of retinal vessels, an U-shaped retinal vessel segmentation algorithm based on the adaptive morphological structure and scale information is proposed. First, the gray image of retina is obtained by synthetically analyzing the three-channel frequency information of the image with two-dimensional K-L (Karhunen-Loeve) transform, and the contrast information between the vessel and the background is enhanced by multi-scale morphological filtering. Then the preprocessed image is trained end-to-end by using the U-shaped segmentation model, and the data is enhanced by local information entropy sampling. The dense deformable convolution structure of the network coding part captures the multi-scale information and shape structure of the image effectively according to the informations of the upper and lower feature layers, and the pyramid-shaped multi-scale dilated convolution at the bottom enlarges the local receptive field. At the same time, introducing deconvolution layer with attention mechanism in decoding phase, which effectively combines the bottom and top feature mappings, can solve the problems of weight dispersion and image texture loss. Finally, the final segmentation result is obtained by using the SoftMax activation function. This approach achieves average accuracies of 97.48% and 96.83% and specificities of 98.83% and 97.75% on the DRIVE (Digital Retinal Images for Vessel Extraction) and STARE (Structured Analysis of the Retina) datasets respectively, which is better than the existing algorithms.
梁礼明, 盛校棋, 蓝智敏, 杨国亮, 陈新建. 自适应尺度信息的U型视网膜血管分割算法[J]. 光学学报, 2019, 39(8): 0810004. Liming Liang, Xiaoqi Sheng, Zhimin Lan, Guoliang Yang, Xinjian Chen. U-Shaped Retinal Vessel Segmentation Algorithm Based on Adaptive Scale Information[J]. Acta Optica Sinica, 2019, 39(8): 0810004.