The details and reproducibility of this method have been described in several publications [4–6]. Briefly, the GSM of blood (B) and adventitia (A) were obtained using the “histogram” facility of the program. This was achieved by selecting an area of noiseless blood from the vessel lumen and the inner two fourths of the brightest area of adventitia adjacent to the plaque (Fig. 8.2). Zooming so that the area of adventitia was enlarged made this procedure easier to perform. Also, selecting the inner two fourths of the brightest area of adventitia was essential for ensuring high reproducibility. Normalization was subsequently performed using the “curves” facility and adjusting the straight line of the “curves” diagram so that the value of B would become zero and the value A would become 190 in the final image. Thus, all the pixels in the image would adjust automatically according to the new linear scale defined by these two reference points. Subsequently, the plaque in the final image was outlined with the mouse and the GSM was obtained from the “histogram” facility.

The use of Adobe Photoshop™ was adequate for image normalization and measurements of GSM as a measure of plaque overall density although time-consuming. However, this software could not provide any measurements of texture, i.e., the spatial distribution of pixel gray scale in the plaque area. This has now been overcome by dedicated software [8].

The Plaque Texture Analysis software LifeQ Medical (info@lifeqmedical.com) which is a dedicated research software package for image normalization and extraction of plaque texture features including GSM became available in 2004. This software has five modules [8]:

1. Image histogram normalization: it provides a user-friendly way to normalize images with blood and adventitia as reference points (Fig. 8.3).

2. Measurements: it provides a means of distance calibration, providing measurements of distance or area in mm and mm², respectively (distances such as IMT, plaque thickness, and areas).

3. Pixel density standardization: it provides a method of normalizing images to a standard pixel density (20 pixels per mm). This is because a number of texture features are pixel density dependent [8]. Pixel density of images from different duplex scanning equipment has been found to vary from 10 to 30 pixels per mm. Also, various degrees of image magnification applied by the operator do alter the pixel density. Thus, the value of 20 pixels per mm has been suggested for a standard image.

4. Image crop: This module has two windows, one for the normalized black and white image and the other for the color-flow image or image with the plaque outlined by the ultrasonographer (Fig. 8.4). The plaque in the normalized image is outlined with the mouse and saved as a new file with the same name and extension “.plq.” Both components of a plaque (anterior and posterior wall) can be selected (Fig. 8.5). By pressing the “Features Extraction” button in this window or using the “Feature Extraction” module, a variety of texture features are automatically calculated (Fig. 8.6) including GSM and can be saved in a database that can be opened by “Windows Excel.” Data can then be transferred to SPSS or any other statistical package.

5. Texture feature extraction: it extracts a number of plaque texture features including GSM (Fig. 8.6) and saves them on a file for subsequent statistical analysis (see section under “texture features” below). It classifies plaques according to the Geroulakos classification [9]. In addition, images of plaques are color contoured: pixels with a gray-scale value in the range of 0–25 are colored black. Pixels with values 26–50, 51–75, 76–100, 101–125, and values greater than 125 are colored blue, green, yellow, orange, and red, respectively. In addition, this module allows printing of the plaque images and selected features in the form of a report or saving the latter in a folder (Fig. 8.7).

For the purpose of automatic classification by computer, the Geroulakos classification has been redefined in terms of pixels and gray levels. Examples of plaque types 1–4/5 are shown in Fig. 8.8. For plaque type 5, only the calcified or visible bright areas of the plaque are selected ignoring the areas of acoustic shadows where information on plaque texture is lacking (Fig. 8.9).

In a comparison and reproducibility study, two operators normalized and measured the GSM of the 33 plaques using both methods [8]. Each image and plaque was initially processed using Adobe Photoshop™ and subsequently using the “Plaque Texture Analysis” software before proceeding to the next image. This ensured that the same area of adventitia and plaque was outlined when using each type of software. The two observers did not know the results of each other. For each observer the GSM values obtained with each software package were compared. The GSM values obtained by each operator using the “Plaque Texture Analysis” software were also compared.

There was a linear relationship between the GSM obtained using the Adobe Photoshop™ program and the Plaque Texture Analysis program with a correlation coefficient (r) of 0.990 (95% CI 0.984–0.996) for the first observer and 0.987 (95% CI 0.973–0.994) for the second observer. The interobserver GSM reproducibility using the Plaque Texture analysis by the two operators had a correlation coefficient (r) of 0.933 (95% CI 0.864–0.967) (p < 0.0001).

The results indicated a high intraobserver reproducibility of GSM when the same plaque images were analyzed using both Adobe Photoshop™ and the dedicated software. In addition, there was a high interobserver reproducibility when the dedicated software was used by each observer.

The effect of image normalization on plaque classification and risk of ipsilateral ischemic neurological events in patients with asymptomatic carotid stenosis was tested in the first 1115 patients recruited to the ACSRS study with a follow-up of 6–84 months (mean 42) [9]. Duplex scanning was used for grading the degree of internal carotid stenosis and for plaque characterization visually (types 1–5) that was performed before and after image normalization and by the “Plaque Texture Analysis” software.

Images that were recorded on video tapes (S-VHS) were digitized off-line on a PC using a video grabber card (Videologic, TV Snap version 1.0.3 c 1990–1994) at a resolution of 640 × 480 pixels at the coordinating center by two members of the team who were experienced in carotid scanning. These two members performed plaque classification. Image normalization was performed by the same members of the team several months later using linear scaling with blood (gray-scale value assigned, 0) and adventitia (gray-scale value assigned, 190) as reference points using Adobe Photoshop™, and the plaques were reclassified without access to the results of the initial classification. As indicated above, before image normalization, plaques with a calcified cap that had more than 15% of the plaque obscured by an acoustic shadow were classified as type 5. After image normalization, both the calcified area and the area of the plaque adjacent to the calcification that was outside the acoustic shadow were considered.

The relationship between plaque classification before image normalization and after image normalization is shown in Table 8.1. Before image normalization 131 plaques were classified as type 1, 288 as type 2, 319 as type 3, 166 as type 4, and 188 as type 5. It can be seen that after image normalization, 66% of type 1, 49% of type 2, 46% of type 3, 66% of type 4, and 82% of type 5 were reclassified as a different plaque type (kappa statistic 0.22) [9].

The ipsilateral neurologic events (AF, TIAs, and stroke) that occurred during follow-up in patients with different types of plaque before and after image normalization are shown in Tables 8.2 and 8.3, respectively. It can be seen that after image normalization the incidence of events in relation to different plaque types has changed. After image normalization there was a decreased incidence in patients with plaques type 4 and 5 with the vast majority of events occurring in plaque types 1, 2, and 3. Before image normalization only 82 (71%) of the 116 neurologic events occurred in plaque types 1–3, but after image normalization, the number increased to 109 (94%).

When plaque types 1–3 were compared with plaque types 4 and 5 before image normalization, the relative risk of having an event was 1.12 (95% CI 0.76–1.66) (Chi Sq. p = 0.45). Also, 37 (73%) of the 51 ischemic strokes occurred in patients with plaque types 1–3 (Table 8.2). When plaque types 1–3 were compared with plaque types 4 and 5 after image normalization, the relative risk of having an event was 4.8 (95% CI 2.27–10.28) (Chi Sq. p = 0.0001). Also, 49 (96%) of the 51 ischemic strokes occurred in patients with plaque types 1–3 (Table 8.3).

When hypoechoic plaques (type 1 and 2) were compared with hyperechoic ones (type 3 and 4), the incidence of ipsilateral stroke was 45 out of 426 (annual stroke rate 3.0%) in the former and 14 out of 635 (annual stroke rate 0.6%) in the latter group (Table 8.4) (Chi Sq. p = 0.003).