基于外轮廓识别与内部像素分类的双阶段血管内超声影像分割与术中显示

Two-Stage Segmentation and Intraoperative Visualization of Intravascular Ultrasound Images Based on Contour Recognition and Internal Pixel Classification

【目的】血管内超声成像常被用于术中观察冠状动脉的狭窄状况与粥样硬化病灶危险程度。本文研发了针对血管内超声影像的病灶智能分割算法,重点解决噪声和伪影对分割精度的影响,并提出便于医生在术中观察的分割结果显示方式。【方法】本文提出了双阶段分割模型,首先自动识别血管内腔边轮廓以排除血管外噪声干扰,然后聚焦于血管内病灶组织的像素分类。本方法充分考虑了时序图像序列在时间上的关联性,通过多通道输入的U-Net提升相邻帧之间的分割一致性。为便于术中观察,根据网络输出的像素概率进行了概率化显示。【结果】在20例时序影像测试集上,本方法对纤维、钙化、脂质与超声衰减分割的平均Dice系数指标分别为0.90、0.93、0.80和0.95。对比实验证明外轮廓识别有助于排除外部噪声干扰,提升内部病灶分割的完整性。本方法以多通道方式输入时序图像可以有效提高时间维度上的分割一致性。临床医生验证肯定了本方法的概率化显示方式有助于术中直观了解病灶分布状况。【结论】本方法通过对血管内超声图像进行分割和可视化展示,更准确、完整、直观地评估并展现血管内病灶成份分布情况,为冠脉介入手术提供了智能分析和可视化展示的支持。

[Objective]Intravascular ultrasound imaging is often used for intraoperative observation of coronary artery stenosis and atherosclerotic plaque risk.In this paper,we develop an intelligent segmentation algorithm for intravascular ultrasound imaging,which focuses on solving the effects of noise and artifacts on the segmentation accuracy,and proposes a way to display the segmentation results that is easy for doctors to observe during the operation.[Methods]In this paper,a two-stage segmentation model is proposed,which first automatically identifies the luminal edge contour of the intravascular vessel to exclude the extravascular noise interference,and then focuses on the pixel classification of the intravascular plaque tissue.This method fully considers the temporal correlation of time-sequenced image sequences and enhances the segmentation consistency between neighboring frames by U-Net with multi-channel input.For intraoperative observation,a probabilistic display is performed based on the pixel probabilities output by the network.[Results]On a test dataset of 20-case time-series images,the average Dice coefficient indexes of fiber,calcification,lipid,and ultrasound attenuation segmentation by the present method are 0.90,0.93,0.80,and 0.95,respectively.Comparison experiments prove that external contour recognition helps to exclude external noise interference and enhance the completeness of the segmentation of internal lesions.This method can effectively improve the segmentation consistency in the time dimension by inputting time series images in a multi-channel manner.Clinicians have confirmed that the probabilistic display of this method is helpful for intraoperative visualization of lesion distribution.[Conclusions]By segmenting and visualizing intravascular ultrasound images,this method can more accurately,completely,and intuitively assess and display the distribution of intravascular lesion components,and provide intelligent analysis and visualization support for coronary interventional procedures.