摘要
目的视网膜光学相干断层成像(Optical Coherence Tomography,OCT)图像黄斑水肿(Macular Edema,ME)区域的自动分割可辅助临床诊断和决策。以提高OCT图像中ME区域分割精度和效率为目的,本文提出了一种联合Dilated U-net网络和全连接条件随机场(Conditional Random Field,CRF)的ME区域全自动分割模型。方法该模型包括两个方面。鉴于U-net网络结构简便和运算高效等特性,用孔洞卷积代替传统卷积增大网络感受野,构建面向OCT图像ME区域分割的网络架构Dilated U-net,实现视网膜ME区域的粗分割。以ME区域的粗分割为初始轮廓曲线,利用全连接CRF实现视网膜ME区域的高精度优化。结果选用200幅OCT图像进行实验,采用精确率、召回率和Dice相似系数对分割精度进行分析,结果表明,本文模型较C-V和SBG等传统分割模型具有更高的分割精确率和Dice相似系数,分别为95.94%和95.52%;较FCN、PSPNet和Deeplab等网络分割模型具有最高的分割效率,单幅OCT图像中ME区域的分割时间降至0.9 s。结论将Dilated U-net与CRF结合改进的ME区域自动分割模型,不仅可获取ME区域的抽象特征,亦兼顾了图像的上下文信息,使ME区域的分割结果更加准确。本文模型不仅提高了病变区域分割的精度,同时降低了分割耗时,适用于视网膜ME区域的高精准分割。
Objective The automatic segmentation of macular edema(ME) region in optical coherence tomography(OCT) image can assist clinical diagnosis and decision-making. In order to improve the accuracy and efficiency of me region segmentation in OCT image, ME region automatic segmentation model of a combined Dilated U-net network and conditional random field(CRF) were proposed in this paper. Methods The model includes two aspects. In view of the characteristics of simple structure and efficient operation of U-net network, hole convolution was used to replace the traditional convolution to increase the network receptive field,Dilated U-net, a network architecture for ME region segmentation of OCT image was established to realize the coarse segmentation of retinal me region. Based on the coarse segmentation of ME region as the initial contour curve, the full CRF was used to optimize the ME region with high precision. Results 200 OCT images were selected for experiments, and the segmentation accuracy was analyzed through accuracy, recall and Dice similarity coefficient. The results showed that this model had higher segmentation accuracy and Dice similarity coefficient than traditional segmentation models such as C-V and SBG, which were 95.94% and 95.52%respectively;compared with network segmentation models such as FCN, PSPNet and Deeplab, it had the highest segmentation efficiency, and the segmentation time of ME region in a single OCT image was reduced to 0.9 s. Conclusion The improved of ME region automatic segmentation model combining dilated U-net and CRF can not only obtain the abstract features of ME region, but also take into account the context information of the image, so as to make the segmentation result of ME region more accurate. This model not only improves the accuracy of lesion region segmentation, but also reduces the segmentation time. It is suitable for highprecision segmentation of retinal me region.
作者
李净
钟元芾
李晓凯
王振华
LI Jing;ZHONG Yuanfu;LI Xiaokai;WANG Zhenhua(Department of Discipline Inspection and Supervision,Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University,Shanghai 201306,China;Department of Information,Shanghai Ocean University,Shanghai 201306,China)
出处
《中国医疗设备》
2021年第11期46-50,66,共6页
China Medical Devices
关键词
黄斑水肿
神经网络
全连接条件随机场
图像分割
macular edema
neural network
fully connected conditional random field
image segmentation
