A novel automatic algorithm for the segmentation of the lumen of the carotid artery in ultrasound B-mode images

Highlights

• A novel algorithm for the segmentation of the carotid artery boundaries is proposed.

• The proposed algorithm does not require human interaction.

• The identification of the lumen region is achieved with efficiency and robustness.

• Satisfactory segmentations of the lumen can be obtained in cases of atherosclerosis.

Abstract

A novel algorithm is proposed for the segmentation of the lumen and bifurcation boundaries of the carotid artery in B-mode ultrasound images. It uses the image contrast characteristics of the lumen and bifurcation of the carotid artery in relation to other tissues and structures for their identification. The relevant ultrasound data regarding the artery presented in the input image is identified using morphologic operators and processed by an anisotropic diffusion filter for speckle noise removal. The information obtained is then used to define two initial contours, one corresponding to the lumen and the other one regarding the bifurcation boundaries, for the application of the Chan-Vese level set segmentation model. A set of longitudinal ultrasound B-mode grayscale images of the common carotid artery was acquired using a GE Healthcare Vivid-e ultrasound system. The results reveal that the new algorithm is effective and robust, and that its main advantage relies on the automatic identification of the carotid lumen, which overcomes the known limitations of the traditional algorithms.

Introduction

The common carotid artery (CCA) is the artery that supplies the human head, specifically the front part of the brain and neck, with oxygenated blood. Like other arteries, it is known for its paired structure: one for left part (with origin in the aortic arch) and another one for the right part of the human body (with origin in the neck). The CCA has a bifurcation in the neck, originating the internal common carotid artery (ICA) and the external common carotid artery (ECA), specifically in the brachiocephalic trunk, and containing a small thoracic portion (Molinari et al., 2007, Rocha et al., 2005, Rocha et al., 2010, Rocha et al., 2011). When compared to the ECA, the ICA is characterized by a lower resistance to waveforms and its lateral and posterior position. The ECA extends anteriorly and medially to the ICA, supplying the face, scalp, neck and tongue with blood, being also characterized by its high resistance to waveforms. Alike other arteries supplying blood from the heart, the carotid is in risk of developing several diseases, in particular, atherosclerosis, known as the “hardening of the artery” (Rocha et al., 2005, Rocha et al., 2010, Rocha et al., 2011).

Atherosclerosis is an inflammatory disease, dominant in the blood vessels, consequent upon the accumulation of fatty substances, mainly lipoproteins and cholesterol in vessel walls. This accumulation is known as “plaque” and causes the narrowing, i.e., the stenosis, of the vessels, decreasing the blood flow (Finn et al., 2010, Maton et al., 1993). In the case of the carotid artery, the bifurcation and the internal carotid artery are the structures more susceptible to atherosclerosis due to the presence of high hemodynamic forces between them. According to recent studies in clinical manifestations of cardiovascular diseases, atherosclerosis has prevalence in two of three men and one in two women after the age of 40 and is characterized as a potential precursor in approximately 60% of deaths (Jennifer and Kathleen, 2009, Latifoglu et al., 2007).

Non-invasive ultrasound imaging has been widely used in the diagnosis of cardiovascular diseases, especially for the evaluation of the intima-media thickness, by assessing the distance between the lumen of the carotid artery, i.e., where the blood flows, and the inner boundary of the adventitia. This measure, and consequent diagnosis of some cardiovascular diseases, particularly atherosclerosis, is performed acquiring B-mode ultrasound images, which requires the detection in these images of not only the lumen boundaries, but also of the near and far adventitia (Hanbay, 2009, Serna-Morales et al., 2012, Stein, 2008). Therefore, it has been an increasing interest in the automatic and robust segmentation of the adventitia and lumen boundaries in ultrasound B-mode images of the common carotid artery. According to Halenka (1999), the carotid adventitia appears in this type of images as two almost parallel lines, known for their echogenic characteristics and separated by a hypoechogenic region, known as the “double line” pattern.

Ultrasound B-mode imaging is the most widely used technique in image-based cardiovascular diagnosis due to the fact of the carotid being a superficial artery and quite suitable for this type of imaging. However, it presents several difficulties, specifically regarding the segmentation of the structures imaged, due to typical image characteristics, such as low contrast, speckle noise, echo shadows and artifacts, which usually lead to images of very poor quality and require the interaction of an expert user for their analysis (Dutt and GrennLeaf, 1994, Wagner et al., 1983). In ultrasound imaging, the speckle noise is presented in echogenic regions in the form of granular texture, and its intensity is related to the scanned tissue. Some works found in the literature have been using different statistical distributions, like, for example, the Rayleigh and K distributions (Dutt and GrennLeaf, 1994, Molthen et al., 1993, Sarti et al., 2005, Wagner et al., 1983), in order to attenuate this granular speckle noise in non-compressed signals. However, most of the signals actually acquired in ultrasound imaging and analyzed in medical practice are log-compressed signals, which are unsuitable for the application of statistical distributions due to their reduced intensity range. In 2006, Noble and co-workers (Noble & Boukerrouni, 2006) described the success of texture segmentation techniques in the classification of breast masses and of liver and kidney tissues in ultrasound images. However, the segmentation of the common carotid artery tends to be more difficult due to the extremely low perception of this structure in usual ultrasound B-mode images.

Ultrasound imaging represents extreme and complex challenges to the automatic segmentation algorithms, as for the reasons aforementioned, but also for the amount of boundary edges that may be omitted in the acquired images, leading to gaps in the computer identified, i.e., segmented, boundaries. Additionally, the scan device may respond differently to dissimilar anatomical structures and reveal distinguished anatomic features and, due to their shape variability among subjects, a model-based segmentation procedure is not appropriate for their successfully segmentation (Rocha et al., 2005, Rocha et al., 2010, Rocha et al., 2011). Despite these difficulties, as mentioned before, there has been an increasing interest in ultrasound imaging-based medical diagnosis, as a consequence of the technological advances verified in this imaging modality, especially regarding its non-invasive characteristic and affordability.

The segmentation of common arteries in usual ultrasound B-mode images can be achieved by two main steps: (i) the definition of a region of interest (ROI) regarding the carotid artery in the input image; and (ii) the detection of the boundaries of the artery lumen, intima and adventitia in the ROI defined. For this reason, we may consider that these two steps are strong interconnected, since the correct detection of the artery walls is strictly dependent on the suitable definition of the ROI.

The first approaches for the segmentation of carotid boundaries in ultrasound B-mode images occurred between 1992 and 1994 (Chan, 1993, Gariepy et al., 1993, Gustavsson et al., 1994, Selzer et al., 1994, Touboul et al., 1992). At this epoch, the computing and image processing and analysis techniques were not as advanced as today, reason for why all these works include a previous manual segmentation of the carotid boundary in order to achieve the final segmentation that was accomplished based on only local image characteristics: echo and gradient intensities. Despite their importance, these segmentation approaches present disadvantages: firstly, the amount of time required, mainly because of the required initial manual segmentation; secondly, the identification of the carotid boundaries cannot be correctly achieved based on a single image characteristic, since it can be strongly affected by speckle noise, low contrast and discontinuities in the carotid boundaries (Molinari et al., 2007, Rocha et al., 2005, Rocha et al., 2010, Rocha et al., 2011).

In 1996, Kozich proposed a new approach based on the minimization of a cost function by dynamic programming. Unlike the segmentation algorithms proposed by Touboul et al., 1992, Gariepy et al., 1993, Selzer et al., 1994, Kozich, 1996, Abdel-Dayem and El-Sakka, 2005a, Abdel-Dayem and El-Sakka, 2005b, Cheng et al., 2011. The Kozich’s approach integrates multiple image characteristics into a cost function. Thus, not only the echo and the gradient intensities were taken into account in the segmentation, but also a local constraint was included in order to originate smoothed segmented carotid boundaries. Each image characteristic and the constraint addressed in Kozich’s work is represented by a cost term, usually a constant expressing the importance or, in other words, the weight of each image characteristic or constraint in the evaluation of the segmentation contour. Because of these cost terms, when in comparison with the solutions developed in previous studies, the one proposed by Kozich led to more robust segmentations with much lower human intervention and consequently, considerably less time consuming. However, it also presents certain disadvantages: its performance is directly affected by the presence of plaques, and therefore, is unsuitable for the correct detection of the carotid boundaries in patients with atherosclerosis; also, with this type of approached, some human intervention is frequently needed to improve the segmentation obtained especially when the input image is of low quality. Besides, it is also required a demanding search for the optimal weights of each image characteristic and constraint, as they have often to be tuned differently for input images acquired using distinct ultrasonic imaging devices or image acquisition parameters.

More recent studies have attained considerable improvements using deformable models that overcome some of the aforementioned problems. These models are focused on the optimization of a cost function in the search of a compromise among the solutions proposed by Kozich, 1996, Selzer et al., 1994, Gariepy et al., 1993 and Touboul et al. (1992), which typically consists in the definition of the smoothness and continuity of the segmentation contour at adjacent points of the model by means of internal forces with the combination of external forces that attract the contour towards the boundaries of the desired structures. Deformable models can be divided into main two classes: parametric models, comprising the active contours, also known as snakes; and geometrical models, usually known as level set models (Ma et al., 2010, Ma et al., 2011).

There are studies, especially those developed by Cheng et al., 1999, Cheng et al., 2002, Schmidt-Trucksass et al., 2001, Jiaoying et al., 2011, Matsakou et al., 2011, Izquierdo-Zaragoza et al., 2011 and Yang et al. (2011), for the segmentation of the common carotid artery based on active contours. However, the application of these models can also be found in the segmentation of other structures in ultrasound imaging, such as the bladder and liver, as well as in other imaging modalities, like magnetic resonance imaging and computerized tomography (Ma et al., 2010). Despite the success cases, usual parametric snakes are not the best choice for an automatic and accurate segmentation of the carotid artery wall, particularly, because the propagation forces of these models are based on image intensity gradient information, often stuck in regions associated to local minimal solutions, and are not robust to speckle noise, false edges or boundary patches at which the image intensity gradient is extremely low or even null (Cheng et al., 1999, Cheng et al., 2002, Ma et al., 2010, Ma et al., 2011). Also, parametric snakes normally require manual initialization, with the definition of an initial contour closest to the carotid boundaries, and constant human intervention to improve the segmentation results (Cheng et al., 1999, Cheng et al., 2002, Loizou et al., 2007, Ma et al., 2010, Ma et al., 2011).

On the other hand, geometrical models present several advantages in relation to parametric models: simple generalization when are transformed from two spatial dimensions to three or even higher dimensions; easy handling of topological changes in the propagating fronts, as in cases where occur split and merge of contours. There are also several works for the segmentation of the common carotid artery based on geometrical models of which it can be referred those addressed by Petroudi et al., 2011, Petroudi et al., 2010, Cheng et al., 2011, Cheng et al., 2011 developed an algorithm that uses parametric snakes, combined with active contours without edges, also known as the Chan-Vese geometrical deformable model, in order to detect the intima-media boundary of the carotid. On the other hand, Cheng, Ding, and Zhang (2011) accomplished a comparison among three different geometric models in the detection of atherosclerosis carotid plaques: the geodesic active contour (GAC) model, the Chan-Vese model and the localizing region-based active contour model. From this comparison, it was concluded that when the initial contour of the deformable model was close to the boundary of interest, the localizing region-based model was the most efficient. The works of Petroudi et al. (2011) and Cheng, Ding, and Zhang (2011) confirm that geometric models are able to detect the carotid wall boundaries proficiently, not only in images of healthy patients, but also in images of patients with atherosclerotic plaques. The segmentation time required by this type of approach is significantly lower than the one required by the parametric snakes, making it appropriate for medical diagnosis routine.

In this paper, a novel computational algorithm is proposed for the automatic identification of the lumen region and consequent segmentation of the lumen boundaries in longitudinal ultrasound B-mode images of the right carotid artery. The algorithm searches for the hypoechogenic characteristics, i.e., for the image pixels with low contrast, of the lumen region of the CCA in the input image, based on the mean and standard deviation of the pixels’ intensity. Afterwards, the lumen and bifurcation boundaries of the CCA are identified using the Chan-Vese geometrical model. The algorithm is robust to speckle noise, does not require human interaction and adjusts adequate the segmentation contours to the lumen boundaries represented in the input B-mode ultrasound images.

The developed algorithm is described in the next section. Afterwards, experimental results are presented and discussed, including the validation of the results obtained by our algorithm. Finally, the conclusions are presented and future work perspectives are pointed out.