Carotid artery image segmentation using modified spatial fuzzy c-means and ensemble clustering

Abstract

Disease diagnosis based on ultrasound imaging is popular because of its non-invasive nature. However, ultrasound imaging system produces low quality images due to the presence of spackle noise and wave interferences. This shortcoming requires a considerable effort from experts to diagnose a disease from the carotid artery ultrasound images. Image segmentation is one of the techniques, which can help efficiently in diagnosing a disease from the carotid artery ultrasound images. Most of the pixels in an image are highly correlated. Considering the spatial information of surrounding pixels in the process of image segmentation may further improve the results. When data is highly correlated, one pixel may belong to more than one clusters with different degree of membership. In this paper, we present an image segmentation technique namely improved spatial fuzzy c-means and an ensemble clustering approach for carotid artery ultrasound images to identify the presence of plaque. Spatial, wavelets and gray level co-occurrence matrix (GLCM) features are extracted from carotid artery ultrasound images. Redundant and less important features are removed from the features set using genetic search process. Finally, segmentation process is performed on optimal or reduced features. Ensemble clustering with reduced feature set outperforms with respect to segmentation time as well as clustering accuracy. Intima-media thickness (IMT) is measured from the images segmented by the proposed approach. Based on IMT measured values, Multi-Layer Back-Propagation Neural Networks (MLBPNN) is used to classify the images into normal or abnormal. Experimental results show the learning capability of MLBPNN classifier and validate the effectiveness of our proposed technique. The proposed approach of segmentation and classification of carotid artery ultrasound images seems to be very useful for detection of plaque in carotid artery.

Introduction

Image segmentation is a critical image processing task which affects the performance of subsequent processing steps. It is the process of grouping pixels of an image into several homogeneous regions. In recent years, medical image segmentation has gained significant importance. The main task is to partition different tissues/organs in a medical image into non-overlapping regions. As a part of several computer-aided diagnosis systems, being used in routine medical treatment, medical image segmentation has drawn the attention of several researchers. For instance, Chuang et al. [1] have proposed fuzzy c-means (FCM) clustering with spatial information for image segmentation. Since standard FCM is sensitive to noise, therefore Chuang's proposed technique considers the spatial information in clustering process to classify the noisy pixels and results in the form of homogenous clustering. In [2], medical image segmentation is proposed based on the moment and transform based features using incremental supervised neural networks. In this technique, images are segmented by using nine features of moments of gray level histogram (MGH) and nine features of two dimensional continuous wavelet transform (2D-CWT) by utilizing supervised neural networks.

Vasantha et al. [3] have proposed medical image features extraction, selection, and classification scheme. In this technique, features of histogram intensity and gray level co-occurrence matrix (GLCM) are extracted and selected for classification by using greedy stepwise method and genetic algorithm.

Li et al. [4] have suggested medical image segmentation based on the integrating spatial fuzzy clustering with level set methods. They have tried to incorporate the concept of partial differential equations with spatial fuzzy clustering. Yu and Tan [5] have formulated object density based image segmentation. They have utilized marker controlled watershed segmentation technique to identify each object of interest. Improper selection of ROI may result in poor clustering. Cárdenes et al. [6] have developed a segmentation evaluation technique based on multidimensional information extracted from medical image data to assess the performance of image segmentation method.

Vascular plaque, a consequence of atherosclerosis, results in an accumulation of lipids, cholesterol, calcification, and other tissues within the arterial wall. It may reduce the blood flow within the arterial wall. Further, this plaque buildup may block partially or fully blood flow to the brain. Furthermore, if the plaque ruptures, small components may drift through the blood into the brain. Results may be in the form of stroke. Early detection of the plaque inside the artery may prevent from serious strokes.

Currently, carotid angiography is the standard diagnosis technique of detecting carotid artery stenosis and the plaque morphology on the arterial walls. This technique involves injecting an X-ray dye and then carotid artery is examined using X-ray imaging. It is an invasive process and it might be uncomfortable and risky for patients, including allergic reaction to the injected dye, kidney failure, and exposure to X-ray radiation.

On the other hand, ultrasound imaging is non-invasive and it is a popular tool for carotid artery assessment. The major shortcoming of the ultrasound imaging is of its low quality due to presence of spackle noise and wave interferences. It demands considerable efforts from radiologists to extract significant information about the contours and plaque layers existing in the artery. This task may require skilled and experienced radiologist. Further, manual extraction of carotid artery contours generates results which may not be reproducible. Therefore, a computer aided diagnostic technique for segmentation of carotid artery contours is highly desirable to help the radiologists to extract significant information about the plaque and to determine the stage of disease [7], [8].

The main purpose of segmenting the artery images is to diagnose the existence of plaque. Mao et al. [9] have proposed a technique for extraction of carotid artery lumen from ultrasound images using deformable model for approximation of the artery lumen. User might be required to initialize the deformable contour. Abolmaesumi et al. [10] have proposed a scheme for automatic detection of intimal and adventitial layers of carotid artery lumen using snake method. Hovda et al. [11] have proposed a new Doppler based imaging scheme in echocardiography with application in blood/tissue segmentation. This technique employs a likelihood ratio function for the classification of blood and tissue signal.

Hamou and El-Sakka [12] have proposed carotid artery ultrasound image segmentation technique based on Canny edge detector [13]. Their technique requires three parameters: standard deviation of Gaussian smoothing kernel, upper and lower bound thresholds. It is used to mask out the insignificant details from the generated edge map. In addition to this, Abdel-Dayem and Ei-Sakka [14] have offered a method for carotid artery contour extraction. This technique uses a uniform quantizer to segment out the image pixels into three classes. These clusters approximate the area inside artery, arterial wall, and background tissues. Morphological edge detector is used to extract edges among these three clusters. Furthermore, to reduce the spackle noise from the ultrasound images they have incorporated an image pre-processing stage and then a post-processing technique is applied to enhance the extracted contours. However this technique is sensitive to noise and cannot differentiate the relevant object with small intensity variations within the arterial wall.

In [15], [16], the watershed segmentation schemes are proposed to segment out the carotid artery ultrasound images. Watershed segmentation scheme, usually results over segmented images. It highly depends upon the selection of threshold during region merging stage based on the difference of average intensity pixels of neighboring areas.

In [8], authors have proposed a fuzzy region growing technique for carotid artery image segmentation. This scheme uses a fuzzy connectedness map for the image, which is a computationally expensive process.

Classification of atherosclerotic carotid plaques of ultrasound images based on the image texture has been proposed by Christodoulou et al. [17]. In their approach, statistical and gray level dependence matrix features are extracted. A total of 22 features are extracted from artery image. These features are combined and used for classification of atherosclerotic. Keeping in view, the large dimension of feature vector, it may be computationally expensive. Rocha et al. [18] have proposed segmentation technique of carotid artery ultrasound images based on RANSAC and cubic splines. They have tested their approach at 50 B-mode carotid artery ultrasound images.

Golemati et al. [19] have proposed Hough transform based carotid artery image segmentation approach. Their proposed approach is quite interesting. However, Hough transform is applied in their approach which detects lines and circles. But, when the vessels are curvy this method may not perform well. They have tested their approach at 10 B-mode noise free ultrasound images. Additionally, they have performed a small scale evaluation of subjects with atherosclerotic vessels and different levels of stenosis.

Loizou et al. [20] have proposed an integrated approach of artery image segmentation. Their approach using snake based model thus needs manual initialization. Inexperienced user may initialize the snakes at improper position which may produce false results. Ultrasound image statistics, law's texture, and neural network for computer aided carotid atherosclerotic analysis technique have been proposed by Mougiakakou et al. [21]. Their proposed approach is user dependent and may not be able to correctly identify ROI.

In this paper, we aim to overcome the shortcomings of previous approaches, i.e., we have enhanced the clustering performance by exploring spatial correlation information and ensemble clustering technique. IMT is measured from segmented images and MLBPNN is used to classify the normal and abnormal carotid artery ultrasound images.

Abbreviations used in the text are given in Table 1. The proposed approach consists of following four stages. First, preprocessing; second, features extraction and selection; third, image segmentation using ensemble technique; and last one, classification of segmented images into normal and abnormal. The paper is organized as follows. The proposed scheme is described in Section 2. Results and discussion is presented in Section 3 whereas, conclusions and future work is described in Section 4.