有条件生成对抗网络的IVUS图像内膜与中-外膜边界检测

针对血管内超声(IVUS)图像中各类斑块、超声阴影和血管分支等造成内膜(LU)与中-外膜(MA)边界难以准确检测的问题,提出一种结合堆叠沙漏网络(SHGN)和有条件生成对抗网络(C-GAN)的IVUS内膜与中-外膜检测的改进方法。首先根据血管形态特点,使用旋转、缩放和Gamma变换等方法将图像训练集扩充57倍,降低网络训练过拟合风险;然后利用对抗训练思想,构建基于L1、L2重建损失的联合损失函数,学习超声图像与其对应分割图像的映射关系,将IVUS图像分割为3种不同区域:血管外周组织、斑块区域和内腔区域;最后在图像分割结果上,采用阈值处理方法,检测最终的内膜与中-外膜边界。采用国际标准IVUS图像数据集(10位病人435幅)评价所提出的算法。实验量化评价结果为:内膜计算面积交并比(JM) 93%,面积差异百分比(PAD) 3%,Hausdorff距离(HD) 0.19 mm;中-外膜JM 95%,PAD 3%,HD 0.16 mm。这些指标满足临床诊断要求,性能优于现有的、近年较好的9种算法,以及Pix2Pix模型。在临床实践应用分析中,利用南部战区总医院心血管内科所收集的100幅IVUS图像进行检验,取得较好的分割结果。这表明该方法具有较好的跨数据集泛化性能。

Lumen and Media-Adventitia Borders Detection in IVUS Images Using Conditional Generative Adversarial Networks

This paper presented an improved method based on stacked hourglass networks (SHGN) and conditional generative adversarial networks (C-GAN) for detecting lumen (LU) and media-adventitia (MA) borders in intravascular ultrasound (IVUS) images to overcome the problems of interference factors, such as different kinds of plaques, ultrasound shadow and branches of vessels. According to the specific shape of coronary artery, several approaches used for data augmentation were applied to increase the number of IVUS samples in training set, including rotating every other 10 degrees, scaling up or down and implementing gamma transformation, in order to reduce the risk of over-fitting during the training stage. Subsequently, in the spirit of adversarial training, a joint loss function based on L1 or L2 loss was constructed and the networks learned a mapping between input IVUS images and their corresponding ground-truths, to segment the whole image into three regions: non-vessel, plaque and lumen regions. Lastly, LU and MA borders were obtained by a thresholding method. Experiments were performed on an international reference dataset containing 435 IVUS images from 10 patients and a small dataset of 100 frames from department of cardiology in Guangzhou General Hospital of Guangzhou Military Region. Quantitative evaluation results showed that the average Jaccard Measures (JM) of two borders of interest were 93% and 95%, the average percentage of area difference measures (PAD) for both was 3%, and average Hausdorff distance measures (HD) were 0.19 mm and 0.16 mm, respectively. These evaluation metrics could meet clinical demands, and the performance of our method was better than those of algorithms evaluated on the same reference dataset and the Pix2Pix model′s performance. Successful detection results on Guangzhou General Hospital dataset demonstrated that the proposed method has good generalization on cross-dataset.