Keywords

1 Introduction

Over the past decade, a rise of 26.6% of cardiovascular cases is noted. According to the experts of [1], the long-term effects of COVID-19 are likely to influence cardiovascular health, and the global burden of cardiovascular disease is expected to grow exponentially over the next few years. The pathological mechanism of cardiovascular diseases starts with deposition of lipoprotein cholesterol, thickening of vessel walls, increased degrees of vascularization, and formation of atherosclerotic plague on the vessel walls [2]. Figure 1a represents a longitudinal B-mode ultrasound image of the carotid artery, and Fig. 1b depicts a plague region in carotid artery. This plague can break off from the inner walls of the blood vessels and enter the circulation causing arterial obstruction. This stimulates thrombus or vasospasm in heart, brain, kidneys, and lower extremities [3,4,5].

Fig. 1
figure 1

Longitudinal B-mode ultrasound image of carotid artery: a healthy, b plague deposits on the arterial wall

Full size image

Precise imaging of the common carotid artery is important for diagnosis as well as assessing the risk of various vascular diseases. There are various imaging modalities used for this purpose, namely ultrasound, invasive coronary angiography, and Magnetic Resonance Imaging (MRI). Angiography is a routine procedure to locate the plague or affected area, but it does not provide structural information [6]. Ultrasonography is the preferred method of vascular imaging due to its reliability, low cost, non-invasiveness, and better visualization of anatomical structures. Important quantitative information such as lumen area, thickness of the carotid walls, distribution and composition of plaques, and intima-media thickness are available in longitudinal ultrasound images [7]. Therefore, analysis of the arterial wall and plague deposits in coronary artery has significant clinical relevance for assessment and management of cardiovascular diseases. Conventionally, these details are measured from ultrasound images by trained personals in clinical setting which is highly user-dependent, time-consuming, and prone to errors. Due to the advent of automation, several computerized techniques have been developed to reduce the subjectivity and time of analysis while increasing the accuracy and efficiency. A large number of follow-up studies have been done on B-mode ultrasound images to determine the possibilities of carotid atherosclerosis or stenosis [3, 8, 9]. A cohort study in [10] reveals that diabetic patient along with renal dysfunction are more susceptible to carotid stenosis leading to mortality.

Since the ultrasound image acquisition procedure is completely manual, the parameter settings of the machine depend on the subjective judgment of the operator. Thus, the acquisition of good quality ultrasound images is complicated and requires high skilled personal. Regardless of the acquisition protocols, the echogenicity depends on tissue composition, insonation angle, tissue attenuation, and blood flow. This leads to local changes in intensity, contrast, and adds speckle pattern in ultrasound images. Hence, particular attention should be given to suppress the speckle pattern before segmentation. The intention of this article is to give an extensive appraisal on the analysis of carotid artery ultrasound images for disease diagnosis starting from image denoising for speckle reduction and segmentation approaches.

2 Preprocessing

Due to ultrasonic echoes from tissues, ultrasound images are often accompanied by multiplying speckle noise. This falsifies the fine details of the image, making it difficult for the computer system to analyze the information. Therefore, despeckling of ultrasound images is critical for better diagnosis of pathologies. Several speckle reduction techniques have been proposed, especially for ultrasound B-mode images. Table.1 shows some of the speckle reduction techniques considered in this study. Commonly available despeckling filters are mean, Lee, Kuan, Wiener, and Gaussian filters. Speckle reducing anisotropic diffusion (SRAD) filter developed by [11] is a benchmark filter for despeckling [12, 13]. Filtering in wavelet domain has also gained popularity among researchers [12, 14]. Linear scaling filter and local statistical filter that utilizes mean and variance are also applied to despeckle the image [15,16,17]. Whereas, authors of [13] developed an integrated toolbox in MATLAB that includes ten despeckling filters specially utilized for carotid artery ultrasound images. On contrary, multiplicative noise is converted into additive noise in [18] and applied wavelet decomposition. Then, three different filters, namely non-local means (NLM), vectorial total variation (VTV), and block matching and 3D filtering (BM3D) algorithms are applied. Whereas, utilized non-local variational model is used in [19] for despeckling of ultrasound images. Table 1 shows the comparative analysis of the despeckling algorithm that are implemented and the performance metrics used for analysis. 

Table 1 Despeckling techniques implemented for longitudinal B-mode carotid artery ultrasound images
Full size table

3 Segmentation

This section highlights the state-of-the art techniques that facilitate segmentation of carotid artery from longitudinal ultrasound images and are listed in Table 2. Researchers developed different methodologies such as edge detection methods, thresholding methods, contours or snakes, and classifier-based approaches.

Table 2 Segmentation techniques implemented for longitudinal B-mode carotid artery ultrasound images
Full size table

Thresholding

Simple thresholding methods like Otsu thresholding using wind-driven optimization, global thresholding followed by morphological operations are utilized in [23, 25]. In [12], modified Otsu thresholding followed by morphological operations is used to segment upper and lower walls.

Edge detection methods

Usually, the blood tissue interface at the walls of blood vessels gives rise to typical edge patterns. These clearly manifest as borders due to discontinuities in the intensity values and reduced edge reflections. Hence, edge detection methods like Prewitt and Sobel operators followed by Hough transform for segmentation are used [16, 27]. Whereas, gradient filter is applied to detect the edges in image followed by watershed algorithm to segment the lesion in ultrasound images [28]. A drawback of this method is over segmentation leading to false detections which is avoided by cluster analysis.

Snake or contours

A variety of snake models are used by researchers in which some are fully automated and some need human intervention for contour initiation. Snake or contour models are edge-based models, wherein the initial contour attracts toward local maxima in the edge map, where the shape of the snake is governed by energy functions. In [15], the boundaries of blood vessels are identified by Chan-Vese level set algorithm which provides excellent initialization for segmentation. The intima-media complex is segmented using active contours as in [21]. Whereas, authors of [29] defined two masks as initial contours for Chan-Vese segmentation of carotid artery walls as well as its bifurcation. The obtained contours are smoothed by cubic spline interpolation and projecting contour points toward local regression line. Active oblongs are utilized in [30] to segment the carotid artery. Hough transform is used to automatically initialize the active oblong in the arterial region followed by pixel offset operations and growing the oblong. The growth of the oblong is optimized using gradient descent technique and Green’s theorem. A similar approach using active oblong with five degrees of freedom is proposed by [31] and added a post-processing step (median filter, the canny edge operator, and the cubic curve fitting) to provide a smooth curved area. In [22], a semi-automatic approach using two frequency implemented B-spline snakes is implemented. In addition, a small gravity force is added to the upper contour, and a take-off force to the lower contour to make sure the contours does not remain motionless. Recently, [26] compared three methods to segment lumen area, namely combining affinity propagation and (DBSCAN) density-based spatial clustering of applications with noise, gradient vector flow (GVF) snake model, and particle swarm optimization clustering-based segmentation. The combination of affinity propagation and DBSCAN outperformed the other models with low computational power. Moreover, the outliers were removed by Z-score method. A limitation of contour-based segmentation is the initialization of contour or snakes. To overcome this problem [24, 32], proposed an intelligent algorithm that locates arterial wall areas to set the initial contour, thus making the system fully automated. Jin et al. [33] proposed fully automated algorithm for region identification, contour initialization, and segmentation of intima-lumen and media-adventia layers using general snake and GVF snake. Authors of [34] proposed H∞ grayscale-derivative constraint snake algorithm to segment intima-media borders and compared with the following models: Kalman snake method, snake method, dynamic programming, and level set method using Chan-Vese energy functional. The layers segmented by H∞ algorithm are precisely defined and robust to system error. The performance of four different snake models (Williams and Shah, Lai and Chin, Balloon, and GVF snake) with manual segmentation is compared [20]. Lai and Chin snake model depicted slightly better region of curve than the other models.

Nakagami models

Nakagami distributions are considerably used in image processing applications as it provides information about the spatial arrangement and statistical distribution of ultrasound imaging data [35, 36]. An iterative method using mixtures of Nakagami distributions and stochastic optimization is applied in [37]. Moreover in [38], they applied Bayesian segmentation approach modeled by mixtures of three Nakagami distributions. Authors have stated that this kind of segmentation approach is semi-computerized and no longer sensitive to the degree of stenosis or calcification. In addition, these images are not preprocessed due to the fact that the application of filter affects the statistics of data.

Dynamic models

Some authors proposed dynamic approach in segmentation. In [39], a combination of dual line detection with edge maps of two edge detectors and coupled snake model is used to maintain parallelism in the intima-media segmentation. Whereas in [40], best fit for cubic spline is searched to identify the adventitia layer. Dynamic programming is employed to segment the lumen boundary. Since dynamic programming causes irregularities in boundaries, a combination of smooth intensity thresholding and hybrid Chan-Vese model is applied as post-processing step.

Other approaches

Simple iterative clustering algorithm was used in [41] to produce super pixels of ultrasound image and RealAdaBoost is utilized to produce classification map. A learning-based algorithm extracted the plague region from the classification map. They segmented the plague region using classifiers such as linear support vector machine (SVM), AdaBoost, random forest, and SVM + radial bias function by considering each pixel in the image as features belonging to either normal or plague area. Additionally, auto-context model is implemented with ten iterations. It was found that random forest is superior, and the context features help to stabilize the model. Whereas, [42] proposed multiclass framework using k-means classifier and proved its superior performance with manual tracings. Conversely, convolutional neural network is used to characterize the plague composition using patch-based approach [43]. In paper [44], an integrated graph model and Markov random fields is used to segment the plague region in coronary artery. Whereas, researchers in [45] suggested segmentation of coronary artery using clustering algorithms. They have utilized fuzzy c-means clustering, spatial fuzzy c-means, modified spatial fuzzy c-means, k-means clustering, and self-organizing maps. K-means and self-organizing maps yielded similar results. Furthermore, equal weightage problem has been overcome by modified spatial fuzzy c-means segmentation. Authors of [42] combined scale space paradigm with a boundary-based approach using level sets, while some used level set method without re-initialization with uniform length [46].

The performance metrics used for comparison of segmentation techniques are as follows: mean, max, min, error, relative error, standard deviation (SD), mean absolute distance (MAD), mean absolute error (MAE), Hausdorff distance (HD), point-to-point distance (PPD), polyline distance (PD), Jaccard index (JI), dice index (DI), Rand index (RI), variation of information (VoI), Cohens kappa (Ck), cophenet (C), sensitivity (Se), specificity (Sp), DSC overlap, EArea, AEArea, point-to-point distance (PPD), Daviea Bouldin index (DBI), partition coefficient (PC), classification entropy (CE), accuracy (Acc), localization (L), true-positive fraction (TPF), and false-positive fraction (FPF), true-negative fraction (TNF), false-negative fraction (FNF), similarity kappa index (KI), and the overlap index (OI), confidence interval (CI), correlation coefficient (CC), coefficient of variation (CoV), percent statistics (PS), goodness-of-fit (GoF), mean contour distance (MCD), computation time (Ct), region of curve (RoC), precision of merit (PoM).

4 Discussion and Conclusion

Several advancements in the field of ultrasonic imaging have been proposed; however, a number of factors hinder automated analysis and disease diagnosis. On analyzing the despeckling approaches, it is evident that some approaches suffer from degraded spatial resolution and increased system complexity. Some techniques face limitations as follows: the window size of filters affects the resultant image quality; hence, a fair choice of window size is necessary; inability to preserve the edges may lead to loss of information; post-processing becomes necessary in some cases.

Whereas, segmentation techniques also had some complications. Boundary-based methods are prone to errors as they are sensitive to gradient variations at the edges [42]. Some authors assumed the arterial structures as straight lines, which is not true in all cases. This has led to false estimation in some scenarios. It is observed that most authors are interested in utilizing contours or snakes for segmentation, but still, they possess some drawbacks such as vulnerable to discontinuities and false edges. Some integrated and dynamic approaches were proposed to overcome these drawbacks. After segmentation, most articles had calculated the characteristics of plague region or thickness of boundary walls which helps in further analysis.

A plethora of techniques has been discussed in this paper. Although better results have been reported, a direct comparison cannot be made due to the following reasons (i) image set used is not same; (ii) some required manual intervention which may introduce human specific errors in final outcome, (iii) different areas of carotid artery (far-end wall, near-end wall, plague, intima-media complex, lumen boundary) are segmented; (iv) some approaches are user-independent, and others were semi-automated. In this context, it is obvious that there is still room for improvement in segmenting carotid artery for real-time diagnosis. Some studies have disclosed that these pathologies may manifest either chronically or acutely in all arterial territories. Hence, a growing interest in vascular imaging and computational analysis helps to expand the diagnostic abilities. Articles like [47, 48] reveal smaller artery size is reported in India, as compared to other western regions of the world due to genetics and lifestyle. Some interpret smaller diameter of arteries in women than men. These conditions are not taken into account in any of the methods. Hence, future works in automated analysis of carotid artery should consider these anatomical differences for better results.