Segmentation method of intravascular ultrasound images of human coronary arteries

Abstract

The goal of this study was to show the feasibility of a 2D segmentation fast-marching method (FMM) in the context of intravascular ultrasound (IVUS) imaging of coronary arteries. The original FMM speed function combines gradient-based contour information and region information, that is the gray level probability density functions of the vessel structures, that takes into account the variability in appearance of the tissues and the lumen in IVUS images acquired at 40 MHz. Experimental results on 38 in vivo IVUS sequences yielded mean point-to-point distances between detected vessel wall boundaries and manual validation contours below 0.11 mm, and Hausdorff distances below 0.33 mm, as evaluated on 3207 images. The proposed method proved to be robust in taking into account various artifacts in ultrasound images: partial shadowing due to calcium inclusions within the plaque, side branches adjacent to the main artery to segment, the presence of a stent, injection of contrast agent or dissection, as tested on 209 images presenting such artifacts.

Introduction

Intravascular ultrasound (IVUS) is a medical imaging modality that produces a sequence of cross-sectional frames of the vascular wall of arteries as a catheter is pulled-back inside blood vessels. It has become very useful for studying atherosclerotic diseases [1].

Various segmentation techniques have been developed for IVUS images of coronary arteries. Among methods that are based on statistics of the B-mode image, a Maximum A Posteriori (MAP) estimator was derived using Rayleigh statistics of the signal for the lumen contour segmentation [2]. A multi-surface 3D graph search using Rayleigh distributions and Chan-Vese terms is proposed in [3]. In [4], a knowledge-based approach is used to determine which level of gray corresponds statistically to the different regions of interest, i.e., the intima, plaque and lumen, in the context of the arterial wall segmentation. In [5], a non-parametric probabilistic model is integrated into a shape-driven method for the segmentation of the arterial wall. Among other probabilistic segmentation methods of the luminal borders, let us mention [6], [7], [8].

Several other techniques than the ones based on speckle statistics have also been proposed in the past five years. Edge information alone [9] or combined gray level intensity attributes [10], [11] were used. Other segmentation algorithms are based on different textural features [12], [13], [14], [15]. Gray level intensity and textural information were combined to detect the lumen boundary [6]; edge attributes were added for the external elastic membrane (EEM) [16]. In these methods, different frameworks were used to extract the vessel wall boundaries from the different image information, classifiers being the most prevalent recently [6], [13], [15], [16]. Threshold and contour filtering [12] and binary morphological operations [14] were also proposed. Taki et al. [9] used deformable models while [10] combined them to graph search. Finally, a multi-agent segmentation [11] and a 3D parallel segmentation method [17] were investigated. Less recent techniques can be found in Section 4.

The aim of this work was to show that an adaptation of the fast-marching segmentation method developed in [18], [19] for femoral artery IVUS segmentation is also powerful in the context of coronary imaging. Compared to femoral arteries [18], [19], the artery wall of the coronaries presents a much more complex movement (see Fig. 1, left image). Moreover, since the IVUS images were acquired at a higher frequency (40 MHz) than in the case of femoral arteries (20 MHz), the lumen presents more speckle and hence, its appearance is more variable in the present study than in the context of [18], [19] (see Fig. 1, right image). Therefore, we considered an original modification of the speed function proposed in [18], [19]. A textural gradient, defined in terms of the distribution of the gray levels in the different components of the vessel wall, was introduced in this new speed function. Moreover, mixtures of gamma probability density functions (PDFs) were used to model the gray level distribution of the log-compressed and filtered envelop of the IVUS images [20], [21] that could not be modeled with Rayleigh distributions assuming uniform scattering tissues [22], as was used in our previous work. We also included a process that computes adaptive weights to calibrate the two components of the speed function to accommodate for various ranges of values coming from different image features. The vessel wall boundaries were modeled as layered contours that propagate simultaneously under that new speed function, which is based on a combination of complementary contour and region information. The multiple interfaces were propagated in the IVUS series of images after having been initially positioned using approximate manual segmentations on 2 perpendicular longitudinal views (L-views) of the 3D volume, to allow the tracking of the artery wall. This type of initialization is adapted for the segmentation of large pullbacks (several millimeters) where the vessel wall components might change across the sequence due to the heterogeneity of the image, and when adjacent cross-sections are discontinuous due to the beating heart movement for acquisitions that are not gated. This segmentation model handles contour irregularities, partial shadowing due to calcium inclusions within the plaque, side branches adjacent to the main artery to segment, the presence of a stent, injection of contrast agent or dissection as often observed for atherosclerotic coronary plaques.