基于流形学习的血管内超声图像序列关键帧的提取及应用

Key frames extraction and application in intravascular ultrasound pullback sequences based on manifold learning

目的以流形学习为基础,提出一种基于血管内超声(IVUS)图像序列的关键帧门控方法,抑制IVUS序列纵切方向上的运动伪影。方法应用流形学习方法中的拉普拉斯特征映射算法,将高维IVUS图像序列降到低维流形中,利用低维特征向量,构建一个距离函数来反映心脏运动规律,将IVUS图像分为心脏舒张末期和非心脏舒张末期两类,从而提取关键帧,组成门控序列。结果临床采集13组IVUS序列(图像915±142帧,血管长度15.24±2.37 mm),计算门控前后图像序列的血管容积、管腔容积和平均斑块负荷。统计实验结果,表明门控序列血管容积、管腔容积显著小于原始序列,门控前后序列的平均斑块负荷差异性不显著,满足临床诊断要求。血管面积方差和管腔面积的方差显著小于原始序列,表明门控序列较原始序列稳定。在IVUS图像序列的纵切图像上,门控序列减少了锯齿形状的运动伪影,与原始序列形状一致,且具有良好的连续性。并将本文方法与已有的提取门控序列方法进行对比。结论本文方法算法简单稳定,抑制了IVUS图像序列的纵向运动伪影。

Objective We propose an image-based key frames gating method for intravascular ultrasound （IVUS） sequence based on manifold learning to reduce motion artifacts in IVUS longitudinal cuts. Methods We achieved the gating with Laplacian eigenmaps, a manifold learning technique, to determine the low-dimensional manifold embedded in the high-dimensional image space. A distance function was constructed by the low-dimensional feature vectors to reflect the heart movement. The IVUS images were classified as end-diastolic and non-end-diastolic based on the distance function, and the IVUS images collected in end-diastolic stage constitutes the key frames gating sequences. Result We tested the algorithm on 13 in vivo clinical IVUS sequences （images 915 ± 142 frames, coronary segments length 15.24 ± 2.37 mm） to calculate the vessel volume, lumen volume, and the mean plaque burden of the original and gated sequences. Statistical results showed that both the vessel volume and lumen volume measured from the gated sequences were significantly smaller than the original ones, indicating that the gated sequences were more stable;the mean plaque burden was comparable between the original and gated sequences to meet the need in clinical diagnosis and treatment. In the longitudinal views, the gated sequences had less saw tooth shape than the original ones with a similar trend and a good continuity. We also compared our method with an existing gating method. Conclusion The proposed algorithm is simple and robust, and the gating sequences can effectively reduce motion artifacts in IVUS longitudinal cuts.