Introduction

Vascular disease, stroke, and arterial dissection or rupture of coronary arteries are considered some of the main causes of mortality in present days. The behavior of the atherosclerotic lesions depends not only on the degree of lumen narrowing but also on the histological composition that causes that narrowing. Therefore, tissue characterization is a fundamental tool for studying and diagnosing the pathologies and lesions associated to the vascular tree.

Although important, tissue characterization is an arduous task that requires manual identification by specialists of the tissues and proper tissue visualization. Intravascular ultrasound (IVUS) imaging is a well suited visualization technique for such task as it provides a cross-sectional cut of the coronary vessel, unveiling its histological properties and tissue organization.

IVUS is a widespread technique accepted in clinical practice to fill up the lack of information provided by classical coronary angiography on vessel morphology. It has a prominent role evaluating the artery lesion after a interventional coronary procedure such as balloon dilation of the vessel, stent implantation, laser angioplasty, or atherectomy.

IVUS displays the morphology and histological properties of a cross section of a vessel [1]. Figure 2.1 shows a good example of different tissues in an

1 Computer Vision Center, Universitat Autonoma de Barcelona, Campus UAB, Edifici O, 08193 Bellaterra (Barcelona), Spain

Soft

Figure 2.1: Typical IVUS image presenting different kind of tissues.

ous plaque

Figure 2.1: Typical IVUS image presenting different kind of tissues.

IVUS image. It is generally accepted that the different kind of plaque tissues distinguishable in IVUS images is threefold: Calcium formation is characterized by a very high echoreflectivity and absorbtion of the emitted pulse from the transducer. This behavior produces a deep shadowing effect behind calcium plaques. In the figure, calcium formation can be seen at three o'clock and from five to seven o'clock. Fibrous plaque has medium echoreflectivity resembling that of the adventitia. This tissue has a good transmission coefficient allowing the pulse to travel through the tissue, and therefore, providing a wider range of visualization. This kind of tissue can be observed from three o'clock to five o'clock. Soft plaque or Fibro-Fatty plaque is the less echoreflective of the three kind of tissues. It also has good transmission coefficient allowing to see what is behind this kind of plaque. Observing the figure, a soft plaque configuration is displayed from seven o'clock to three o'clock.

Because of time consumption and subjectivity of the classification depending on the specialist, there is a crescent interest of the medical community in developing automatic tissue characterization procedures. This is accentuated because the procedure for tissue classification by physicians implies the manual analysis of IVUS images.

The problem of automatic tissue characterization has been widely studied in different medical fields. The unreliability of gray-level only methods to achieve good discrimination among the different kind of tissues forces us to use more complex measures, usually based on texture analysis. Texture analysis has played a prominent role in computer vision to solve tissue characterization problems in medical imaging [2-9].

Several researching groups have reported different approximations to characterize the tissue of IVUS image.

Vandenberg in [10] base their contribution on reducing the noise of the image to have a clear representation of the tissue. The noise reduction is achieved by averaging sets of images when the least variance in diameter of the IVUS occurs. At the end, a fuzzy logic based expert is set to discriminate among the tissues.

Nailon and McLaughlin devote several efforts to IVUS tissue characterization. In [11] they use classic Haralick texture statistics to discriminate among tissues. In [12] the authors propose the use of co-occurrence matrices texture analysis and fractal texture analysis to characterize intravascular tissue. Thirteen features plus fractal dimension derived from Brownian motion are used for this task. The conclusion shows that fractal dimension is unable to discriminate between calcium and fibrous plaque but helps in fibrous versus lipidic plaque. On the other hand, co-occurrence matrices are well suited for the overall classification. In [13], it is stated that the discriminative power of fractal dimension is poor when trying to separate fibrotic tissue, lipidic tissue, and foam cells. The method used is based on fractal dimension estimation techniques (box-counting, brownian motion, and frequency domain).

Spencer in [14] center their work on spectral analysis. Different features are compared: mean power, maximum power, spectral slope, and 0 Hz interception. The work concludes with the 0 Hz spectral slope as the most discriminative feature.

Dixon in [15] use co-occurrence matrices and discriminant analysis to evaluate the different kind of tissues in IVUS images.

Ahmed and Leyman in [16] use a radial transform and correlation for pattern matching. The features used are higher order statistics such as kurtosis, skew-ness, and up to four order cumulants. The results provided appear to have fairly good visual recognition rate.

The work of de Korte and van der Steen [17] opens a new proposal based on assessing the local strain of the atherosclerotic vessel wall to identify different plaque components. This very promising technique, called elastography, is based on estimating the radial strain by performing cross-correlation analysis on pairs of IVUS at a certain intracoronary pressure.

Probably, one of the most interesting work in this field is the one provided by Zhang and Sonka in [18]. This work is much more complex trying to evaluate the full morphology of the vessel. Detecting the plaque and adventitia borders and characterizing the different kind of tissues, the tissue discrimination is done using a combination of well-known techniques previously reported in the literature, co-occurrence matrices and fractal dimension from brownian motion, and adding two more strategies to the amalgam of features: run-length measures and radial profile. The experiments assess the accuracy of the method quantitatively.

Most of the literature found in the tissue characterization matters use texture features, co-occurrence matrices being the most popular of all feature extractors. Further work has been done trying to use other kind of texture feature extractors and IVUS images, and although not specifically centered on tissue characterization, the usage of different texture features in plaque border assessment is reported, which can be easily extrapolated to tissue characterization. In [19], derivatives of Gaussian; wavelets, co-occurrence matrices, Gabor filters, and accumulation local moments are evaluated and used to classify blood from plaque. The work highlights the discriminative power of co-occurrence matrices derivatives of Gaussian and accumulation local moments. Other works such as [20] provide some hints on how to achieve a fast framework based on local binary patterns and fast high-performance classifiers. This last line of investigation overcomes one of the most significant drawbacks of the texture based tissue characterization systems, the speed. Texture descriptors are inherently slow to be computed. With the proposal of the feature extractor based on local binary patterns a good discriminative power is ensured as well as a fast technique for tissue characterization.

Whatever method we use in the tissue characterization task, we follow an underlying main methodology. First, we need to extract some features describing the tissue variations. This first step is critical since the features chosen have to be able to describe each kind of tissue in a unique way so that it cannot be confused with another one. In this category of feature extraction we should consider the co-occurrence matrix measures, local binary patterns, etc. The second step is the classification of the extracted features. Depending on the complexity of the feature data some methods will fit better than others. In most cases, high-dimensional spaces are generated, so we should consider the use of dimensionality reduction methods such as principal component analysis or Fisher linear discriminant analysis. Either a dimensionality reduction process is needed or not, this step will require a classification procedure. This procedure can be supervised, if we provide samples of each tissue to be classified so that the system "knows" a priori what the tissues are, or unsupervised, if we allow the system to try to find which are the different kind of tissues by itself. In this category, we can find clustering methods for unsupervised classification and, for supervised classification, methods like maximum likelihood, nearest neighbors, etc.

The following sections are devoted to describe the following: First, the most significative texture methods used in the literature. Secondly, some of the most successful classification methods applied to IVUS characterization are reviewed. Third, we describe the result of using such techniques for tissue characterization and conclude about the optimal feature space to describe tissue plaque and the best classifier to discriminate it.

0 0

Post a comment