| Coronary artery heart disease, referred to as coronary heart disease (CHD), is the myocardial dysfunction and/or organic disease due to coronary artery stenosis and insufficient blood supply, which is the common cause of death in the world and known as’the first killer’ and among which 95%～99% are the coronary atherosclerosis. With the improved living standards, changes in diet style and life style, as well as the increase in aging population, the prevalence and mortality rates of coronary heart disease in our country have increased gradually.With the rapid development of modern technology and further study in the coronary heart disease, coronary imaging technology has become an important way to visually observe the anatomy of blood vessels and analyze lesions characteristic during the clinical diagnosis of coronary heart disease, among which coronary angiography (CAG) and intravascular ultrasound imaging (IVUS) are the representatives. Over the past 3 decades, CAG has been supposed as the’golden standard’ to evaluate coronary anatomy. However, it only shows the outline of the lumen of vessels without providing information about the structures of the vascular walls. Pathological studies suggest that there is significant difference between CAG images and the pathology results. Recently, IVUS has been widely used in the diagnosis and therapy of coronary artery disease, particularly in the percutaneous coronary intervention (PCI). It is the combination of non-invasive ultrasound technology and invasive catheter technology with miniaturized ultrasound transducer placed within coronary artery through a catheter during the imaging process. Morphological and pathological characteristics of the vascular cross-section were shown in the IVUS images according to acoustic characteristics of tissues. IVUS has been considered to be the new’golden standard’in the diagnosis of coronary heart disease because it is used not only to accurately measure lumen diameters, areas and other indicators, but also to analyze plaque components according to acoustic characteristics.External elastic membrane (EEM) analysis of the IVUS images is not only an important measurement of atherosclerotic plaque, but also the premise of coronary artery disease diagnosis and interventional treatment. It is commonly performed by clinicians in clinical practice and the number of frames needed to be segmented can range from a few frames to hundreds frames according to the types of analysis, which means that the manual delineation of EEM may be excessively time-consuming and suffer from high level of subjectivity. Therefore, developing computer-aided algorithms for automatically segmenting IVUS images is important to improve efficiency and quality of the clinical diagnosis and treatment of coronary heart disease.This article is aimed to automatically detect the EEM contours in IVUS images to assist in the diagnosis and therapy of coronary artery disease. Due to the interferences such as guide wire, calcified plaque, stents, as well as image artifacts and noises, most existing EEM detection algorithms could not be widely applied in clinical practice. Edge detectors such as Sobel, Canny and DoG used in the early studies were susceptible to the stronger responses of stents and calcified plaques. Then texture features such as Gabor filters, wavelet decomposition, GLCM were added to improve the accuracy of detection algorithm. In addition, methods based on gray-level probability have attracted extensive attention in recent years. Despite of all, problems caused by acoustic shadow, dense fibrous plaque and other disturbing factors still exist. Manual interactions were suggested in several studies to modify the results misguided by interferences but with the cost of reducing the degree of automation.Although the way of clinicians to manually delineate the EMM contours has been rarely considered in most EEM detection algorithms, methods with computer-aided simulation of artificial vision in the image processing and analysis are proved to achieve better results. During the clinical practice, some’critical points’for EEM contours are firstly identified by clinicians in the recognizable or important regions and the complete contours are subsequently accomplished based on the prior vascular shape information. An improved EEM detection method based on sequential learning and prior shape information is proposed in this article to solve the problems presented above simulating the clinicians’observation of EEM. Three important algorithms in the proposed method are shown in the following.First of all, IVUS images are separated into seven tissues by a supervised learning algorithm utilizing the gray-level, edge and texture features which are extracted by the commonly used feature detectors such as Loggabor filter, Sobel, and LBP. The classification model established by a supervised learning algorithm is applicable to other databases if training samples are sufficient. Moreover, the sequential learning algorithm combined with context information can be adapted to deal with the specific spatial relationships between different tissues efficiently in IVUS images. Apart from these, the identification of plaque, shadow and other disturbing factors contribute to eliminating their effects during the EEM detection.Secondly, initial EEM contours are detected based on prior shape information as follows.’Critical points’can be identified according to classification results with spatial relationships between EEM contour and tissues. In order to reduce influences of unavoidable classification errors, an algorithm is proposed to detect and eliminate the unsatisfied contour segments according to prior vascular cross-sectional shape. And then heuristic graph search with heuristic shape information is used to connect the selected’critical points’forming a smooth closed initial contour.Thirdly, the final EEM contours are detected based on a Snake model where the most important is to design an external energy function reasonably. After initial contours are identified with structure constraints from above, two types of image information related to EEM contours are adopted to construct the external energy function. The first one is the edge information presenting the gray-level transition from low to high near the EEM interface. The second one is the media structure information extracted by phase symmetry. In addition, in order to ensure the detected contours are located below the calcified plaque and go through the acoustic shadow zone, an energy compensation term is introduced to compensate the missing information in the acoustic shadow zone.The dataset for testing and validation consists of 153 representative images which are selected by the clinician from a set of 67 pullbacks acquired from 22 patients in Nanfang Hospital, in which various types of images with plaques, acoustic shadow, guide-wire, eccentric vessel are included and standard EEM contours are provided by clinicians. The common classification algorithm and sequential learning algorithm achieved accuracy of 82.39% and 87.54% respectively through 5-fold cross validation, which means that sequential learning algorithm with the consideration of multiscale spatial information significantly improves the ability to recognize different tissues and the segmentation accuracy of IVUS images. The vascular structures are better defined with less noise in the proposed algorithm. The EEM contours extracted by our algorithm meet the requirement of clinical diagnosis with the average performance measurements shown as follows:JACC=88.5%, HD=0.3755mm, PAD=8.83%. The average Hausdorff distance between the detected initial contours and the standard contours after curvature correction is 0.35mm while the distance is 0.53mm before correction. It means that the initial contours with the constraints of vascular shape are more consistent with the standard. It has been proved that the performance of the algorithm proposed in this article is better than the performance of domestic EEM detection algorithms in recent years. The algorithm in this article is less susceptible to collateral vessels and vascular bifurcation with better ability to identify fibrous plaque. So in conclusion, the EEM automatic detection algorithm proposed in this paper is simple and effective. It improves the ability to identify the calcified, fibrous plaques as well as shadow areas compared to domestic existing algorithms. |
PDF Full Text Request