An IVUS image-based approach for improvement of coronary plaque characterization

Abstract

Virtual Histology-Intravascular Ultrasound (VH-IVUS) is widely used for studying atherosclerosis plaque composition. However, one of the main limitations of the VH-IVUS relates to its dependence to the Electrocardiogram (ECG)-gated acquisition. To overcome this limitation, this paper proposes a robust image-based approach for characterization of the plaques using IVUS images. The proposed method consists of three main steps of (1) shadow detection: as an efficient preprocessing step to identify and remove acoustic shadow regions; (2) feature extraction: a combination of gray-scale based features and textural descriptors; and (3) classification: to classify each pixel into one of the three classes (calcium, necrotic core and fibro-fatty). In order to evaluate the efficiency of the proposed algorithm two in-vivo and ex-vivo data sets are considered. The kappa values of 0.639 on in-vivo and 0.628 on ex-vivo tests with VH-IVUS and the histology images labeled by the experts respectively indicate the effectiveness of the proposed algorithm.

Introduction

Atherosclerotic plaques are known to be the most common cause of cardiovascular diseases [1]. Analyzing atherosclerotic plaque composition is a helpful procedure to diagnose such diseases, since the distribution and morphology of components of these plaques define the severity of the lesion [1], [2].

Intavascular Ultrasound (IVUS) is a catheter-based imaging technique which provides real-time high-resolution images of the vessel wall and lumen [3]. During the last decade, several techniques have been developed for atherosclerotic plaque characterization [4], [5].

The Virtual Histology IVUS (VH-IVUS) [6], [7] is the most prevailing and widely available technique for in-vivo plaque characterization [8], [9], [10] (In fact, [7] has been cited 627 times based on Google Scholar which shows the popularity of this technique). VH-IVUS is based on radiofrequency (RF) signal analysis of reflected ultrasound pulses for plaque characterization which is able to differentiate plaque components of fibrous, fibro-lipid, calcified, and necrotic tissues [6], [7].

In VH-IVUS technique, to minimize both the RF attenuation and shifting due to the presence of blood in coronary arteries, an ECG-gating procedure has been performed prior to plaque characterization. As a consequence, the RF spectrum from only one IVUS frame in each cardiac cycle is recorded which is synchronized with the R-wave in the ECG. More precisely, if we consider an IVUS imaging procedure with the rate of 30 frames/s and the normal heart rate assumed to be 1 beat/s, only one IVUS frame out of 30 is processed, i.e. almost 96% of data is discarded [10], [11]. Hence, ECG-gating decreases dramatically the longitudinal resolution of these methods. One of the objectives of this paper is increasing the number of characterized IVUS frames, which is essential in observing the structure (shape) of the different plaque components in the plaque area, which in turn would lead to an improved diagnosis of the disease.

Image processing based methods, on the other hand, hold the potential to provide objective and quantitative measures of plaque composition. Previous studies have identified texture analysis as being useful in the analysis of ultrasound images. Hence, almost all image-based plaque characterization methods are based on texture analysis. In [12], a method to classify plaque regions into three classes: soft plaque, hard plaque and hard plaque shadow was proposed. Gray-level-based wavelet descriptors, co-occurrence matrices, run-length measures and fractal-based measures were selected as descriptors. This plaque characterization method was limited to classification of soft and hard plaques while methods providing more detailed plaque characterization are desired for a better understanding of plaque composition. It was attempted at classifying plaque regions into finer classes of calcified, fibrous and necrotic core in [13]. First-order statistics, Haralick's feature, and Laws' texture energy features were used. However, their developed method was time-consuming and not suitable for near real-time IVUS image assessment. In [14], a texture-based algorithm for plaque characterization using discrete wavelet packet frame (DWPF) was developed. They showed that the extracted textural features are perfectly suited for classification and capturing characteristics of the plaque with the highest correlation to histology. Additional experiments were performed and reported in [15], [16]. Although this method could discriminate fibrotic tissue from lipid however, it could not detect the necrotic core components directly.

A common shortcoming of the currently available signal and/or image based methods is that the shadow regions are treated in a similar way to the other parts of the image. Acoustic shadowing usually occurs behind the dense calcified (DC) regions making them to be classified as lipid or fibro-fatty (FF) classes while they are mostly calcium and necrotic core (NC) plaques [17], [18]. This important problem was recently tackled in a different way in [19] by some of us. First, a new class of image regions which is called “shadow region” was added to the possible plaque components. Then a post-processing step based on the intensity variations of the image pixels, taking the VH-IVUS images as the reference for comparison, was used [19]. This post-processing step was, however, very sensitive to the threshold values used making the whole method to be data set dependent.

The contribution of the current paper, which can be seen as a continuation of our work in [19], [20], is four-fold; first, the thresholds used in the shadow detection step, which was empirically set in [19], are derived from manually characterized images by the experts. The detected shadow regions are then examined by comparing them to those manually characterized by experts. Second, different feature extraction methods are combined to benefit from their diverse characteristics simultaneously and to dispose the naive post-processing step used in [19]. The two well-known classifiers of Support Vector Machines (SVM) [21] and Error-Correcting Output Codes (ECOC) [22] are then utilized producing robust results. Third, extensive experimental and validation studies, including both in-vivo and ex-vivo datasets with real histology images, are carried out and the performance of the proposed plaque characterization method is reported. Indeed, the performance of an IVUS plaque characterization technique is scrutinized in both in-vivo and ex-vivo datasets. And finally, it is demonstrated that the proposed method provides more information for clinicians by characterizing the IVUS frames missed in between VH-IVUS frames.

The rest of the paper is organized as follows. The proposed algorithm is introduced by its different steps in Section 2. Section 3 provides experimental results of the application of the proposed algorithm to real data. In Section 4, the robustness of the proposed algorithm is studied. More explicitly, the performance of the algorithm on a new in-vivo and an ex-vivo datasets is examined. Section 5 demonstrates how the proposed algorithm would be able to overcome the VH-IVUS limitation. Finally Section 6 discusses the pros and cons of the proposed algorithm and concludes the paper.