Computerized Texture Analysis of Carotid Plaque Ultrasonic Images Can Identify Unstable Plaques Associated With Ipsilateral Neurological Symptoms

Transcription

1 Carotid Artery Disease Computerized Texture Analysis of Carotid Plaque Ultrasonic Images Can Identify Unstable Plaques Associated With Ipsilateral Neurological Symptoms Angiology 62(4) ª The Author(s) 2011 Reprints and permission: sagepub.com/journalspermissions.nav DOI: / Stavros K. Kakkos, MD, MSc, PhD 1, Andrew N. Nicolaides, MS, FRCS 1, Efthyvoulos Kyriacou, PhD 2,3, Stella S. Daskalopoulou, MD, MSc, PhD 1,4, Michael M. Sabetai, MD, PhD, FRCS 1, Constantinos S. Pattichis, PhD 3, George Geroulakos, MD, PhD, FRCS 1, Maura B. Griffin, DMU, MSc, PhD 5, and Dafydd Thomas, MA, MD, FRCP 6 Abstract We estimated the value of objective, computerized texture analysis of ultrasonic images in distinguishing carotid plaques associated with neurological ipsilateral symptoms (amaurosis fugax [AmF; n ¼ 30], transient ischemic attack [TIA; n ¼ 52], and stroke [n ¼ 55]) from asymptomatic plaques (n ¼ 51). We performed 3 case-control studies (1/symptom with asymptomatic plaques as control). On logistic regression, AmF was independently associated with severity of stenosis, percentage of pixels with gray levels 0 to 10 (PPCS1; measure of echolucency), and spatial gray level dependence matrices (SGLDM) information measure of correlation (IMC-1; texture); TIAs with PPCS1 (echolucency), SGLDM correlation, and skewness (both texture); and stroke with PPCS1, SGLDM correlation, and percentage of pixels with gray levels 11 to 20 (PPCS2; echolucency). The area under the curve of the regression-derived predicted probability for AmF, TIA, and stroke was 0.92, 0.82, and 0.85, respectively (all P <.001). Texture analysis can identify carotid plaques associated with a neurological event, improving the diagnostic value of echolucency measures. Texture analyses could be applied to natural history studies. Keywords stroke, carotid artery plaque, ultrasound, texture analysis, echodensity Introduction There is considerable evidence suggesting an association between symptomatology and plaque echodensity, which is a measure of overall plaque brightness. Previous ultrasonic studies, based on subjective visual assessment, demonstrated that symptomatic carotid plaques, especially those producing amaurosis fugax (AmF), appear echolucent, with low grayscale median (GSM), namely the median of gray values of the pixels within the plaque. 1,2 However, there is only limited research on the association between textural features of carotid plaque ultrasonic images and patient symptoms 1-3 ; textural features objectively estimate not only plaque echolucency but importantly also plaque heterogeneity. A GSM cut-off point of 40 was shown to achieve only partial separation of symptomatic from asymptomatic plaques, 4 with an odds ratio of approximately 4; therefore, the additional use of statistically independent textural features might further improve plaque separation and shed light on plaque morphology. Texture analysis has already been shown to help predicting the presence of symptoms, 5 but whether it can predict different kind of hemispheric symptoms has not been tested before. The aims of the present study were to 1) Identify textural features of ultrasonic images of carotid plaques independently associated with AmF, transient ischemic attack (TIA), and stroke, and 1 Department of Vascular Surgery, Imperial College London, London, UK 2 Department of Computer Science and Engineering, Frederick University, Limassol, Cyprus 3 Department of Computer Science, University of Cyprus, Nicosia, Cyprus 4 Department of Medicine, McGill University, Montreal, Quebec, Canada 5 The Vascular Noninvasive Screening and Diagnostic Centre, London, UK 6 Department of Neurology, St Mary s Hospital, London, UK Corresponding Author: George Geroulakos, Department of Surgery, Faculty of Medicine, Imperial College, Charing Cross Hospital, Fulham Palace Road, London W6 8RF, UK g.geroulakos@imperial.ac.uk

2 318 Angiology 62(4) 2) Test the hypothesis that a combination of textural and histogram features (including simple measures of echodensity), when compared with simple measures of plaque echodensity alone can better distinguish asymptomatic plaques from plaques associated with AmF, TIA, and stroke. Materials and Methods A total of 43 patients with 51 asymptomatic hemispheres (including 8 patients with bilateral lesions) and 137 patients (137 hemispheres) with AmF (n ¼ 30), hemispheric TIA (n ¼ 52) or stroke (n ¼ 55) were included in this crosssectional study. These were consecutive patients who were referred to the vascular laboratory and were identified with internal carotid artery stenosis 50% using ultrasonic velocity criteria; stenosis was expressed as a percentage in relation to the carotid bulb according to the European Carotid Surgery Trial (ECST) criteria. 6,7 Symptomatic patients had to have had experienced their event within the preceding 6 months to be included in the study. The most recent event was considered. For example, in patients with multiple TIAs, the most recent was considered (in terms of timing), and patients with a recent stroke who have had previous TIAs were classified in the stroke subgroup. Asymptomatic patients never had symptoms. Patients with potential cardioembolic causes of hemispheric symptoms (including arrhythmia, recent myocardial infarction, and endocarditis) were identified and excluded on clinical grounds. Also excluded were patients with atypical symptoms indicating other pathology, distant symptomatology (>6 months, considered by some as asymptomatic), or mild stenosis (<50%, ECST criteria). To avoid possible systemic effects related to plaque instability, asymptomatic (contralateral) sides of symptomatic patients were not included. 8 To minimize a possible effect of stenosis severity on symptomatology, asymptomatic controls were selected so that their stenosis severity matched the stenosis severity of symptomatic patients. Informed consent was obtained from all participants and the protocol was approved by the institutional review board. Carotid ultrasound scanning and off-line image analysis were performed as previously described. 9 Briefly, ultrasound examinations were performed using duplex scanning and colour flow imaging (ATL HDI 3000, Advanced Technology Laboratories, Bothell, Washington). Duplex scans and plaque images were recorded on S-VHS videotapes. A S-VHS Panasonic video cassette recorder (model AG-7350-B, Matsushita electric Ind Co Ltd, Japan) and an external capturing device-frame grabber (Snappy v2, Play Inc., Ca, USA) were used to transfer the images from the videotapes to a computer, saved as Tag Image File Format (Aldus TIFF) files. As previously suggested by Haralick et al., texture analysis should be performed after standardization of the overall image brightness (normalization). 10 Therefore, following the image digitization process, we performed a manual standardizing procedure developed by our group (brightness normalization, using Adobe Photoshop version 5.5, Adobe Systems Inc, Palo Alto, California). This method uses blood and adventitia as reference points, and it results in a significant improvement in the comparability of the ultrasonic tissue characteristics. 9 Bicubic interpolation (resolution standardization) was subsequently employed to resize all images at a fixed resolution (pixel density) of 20 pixels/mm, before texture analysis was performed, as previously described. 11 The original resolution of the 188 images was ranging between 10 and 56 (median 17.3, interquartile range ). Subsequently, visual plaque classification (subjective method) and computer-assisted plaque analysis (objective method) were performed. The analyst was blinded to the clinical status of the patients. Visual Plaque Classification Plaques were classified into 5 types according to visual estimation of overall echodensity: type 1 (completely echolucent lesions, sometimes a thin fibrous cap is visible), type 2 (predominantly echolucent lesions having less than 50% echogenic components), type 3 (predominately echogenic with less than 50% echolucent components), type 4 (uniformly dense echogenic lesions with less than 10% echolucent components) and type 5 (calcified plaques with excessive acoustic shadow). 12 The presence of a substantial (15% or more of the total plaque area) echolucent plaque component near the lumen (juxtaluminal) 13 and discrete echogenic plaque components were also recorded. Computer-Assisted Plaque Analysis Using commercially available software ( Plaque Texture Analysis software, Iconsoft International Ltd, PO Box 172, Greenford, London UB6 9ZN, UK), carotid plaques were manually outlined, by one of the authors (SKK), with the aid of the corresponding color-coded image. If present, near and far wall plaques were combined and included in the plaque outline. The ROI carotid plaque outline was saved as a separate file, which was used for subsequent analysis. Fifty-one histogram and texture features (Appendix I) of the gray tones produced within the outlined plaque area were extracted using the following methods: (i) Histogram measures 14 (ii) First-order gray level parameters, including GSM 15,16 (iii) The spatial gray level dependence matrices (SGLDM) 10 (iv) Gray level difference statistics (GLDS) 17 (v) Gray level run length statistics 18 (vi) The Fourier power spectrum (FPS) 17 Methods (iii) to (vi) are established texture analysis algorithms used to study heterogeneity. All methods were applied on the ROI selected by the expert physician; pixels used for the analysis were only the pixels of the ROI, and not the pixels of the rectangle that includes the ROI, as several other applications do.

3 Kakkos et al 319 Table 1. Plaque Type in Symptomatic and Asymptomatic Patients Neurological Status Asymptomatic Amaurosis Fugax Transient Ischemic Attack Stroke Plaque type 1 3 (5.9%) 14 (46.7%) 12 (23.1%) 15 (27.3%) 2 21(41.2%) 16 (53.3%) 32 (61.5%) 29 (52.7%) 3 21(41.2%) 0 (0%) 7 (13.5%) 9 (16.4%) 4 6 (11.8%) 0 (0%) 1 (1.9%) 2 (3.6%) 1 and 2 (echolucent) 24 (47.1%) 30 (100%) 44 (84.6%) 44 (80%) 3 and 4 (echogenic) 27(52.9%) 0 (0%) 8 (15.4%) 11 (20%) a Echolucent plaques were more frequently observed in symptomatic patients (P <.001). Table 2. Distribution of Juxtaluminal Echolucent Plaque Components in Symptomatic and Asymptomatic Patients; These Were More Frequent in Symptomatic Patients (P <.001) Neurological Status Asymptomatic Amaurosis Fugax Transient Ischemic Attack Stroke Juxtaluminal echolucent area Absent Present 23 (45%) 28 (93%) 42 (81%) 47 (86%) Odds ratio (95% CI) a 17 (3.7-79) 5.1 ( ) 7.2 ( ) P value a <.001 <.001 <.001 a Asymptomatic plaques versus amaurosis fugax, transient ischemic attack and stroke. Statistical Analysis Chi-square was used to assess the association between symptoms and categorical variables. The Kolmogorov- Smirnov test was used to test data for normal distribution; in case of nonnormally distributed data, nonparametric tests Mann-Whitney and Kruskal-Wallis were used to compare numerical data distribution of 2 and more than 2 groups, respectively. Factor analysis (principal components method with varimax rotation) was used to identify collinearity and exclude redundant histogram and texture variables before performing multivariate analysis. Data were linearly transformed whenever possible before entering them into factor analysis. Multivariate analysis with binary logistic regression (forward and backward Wald methods) was used to identify the features (degree of stenosis, plaque area, histogram, and texture measures) that were independent predictors of patient symptoms. Previous studies 1,3-5 have identified different echostructure between symptomatic and asymptomatic plaques. Therefore, to explore whether texture analysis can identify plaques associated with different symptomatology, symptomatic plaques were divided into 3 groups according to the specific symptom (ie, AmF, TIAs, and stroke). Cox & Snell R-square was used to determine the proportion of the variation that is explained by the model, while the Hosmer and Lemeshow test was used to test model goodness-of-fit. For each individual case, the statistical software saves the predicted probability (referred to as model output) of occurrence of the event. The significance of the various texture features and derived scores (including the output of logistic regression, based on the independent variables, known as predicted probability) was assessed with the receiver operating characteristic (ROC) curves method, and expressed as the area under the curve and 95% confidence intervals (CI). 19 The ROC curves analysis was repeated for the 3 comparisons/groups of the study. Receiver operating characteristic curves were compared using Hanley s method 20 and MedCalc for Windows (version 7, Mariakerke, Belgium). The remaining statistical analysis was performed using the SPSS for Windows statistical package (release 11.5, Chicago, Illinois). Results The degree of stenosis (median, interquartile range) was similar between the group of symptomatic patients (90%, 83%-94%) and the group of asymptomatic patients (90%, 80%-94%; P ¼.74, Mann-Whitney test). The degree of stenosis (median, interquartile range) in symptomatic patients with AmF, TIAs, and stroke was 94% (83%-94%), 90% (81%-94%), and 90% (80%-94%), respectively (P ¼.41, Kruskal-Wallis test). Symptomatic patients had their ultrasound examination performed within a month from their clinical event. Visual Plaque Classification The association between plaque types and patient symptoms is shown in Table 1. Plaque types 1 and 2 (echolucent) were

4 320 Angiology 62(4) Table 3. Independent Parameters Associated With Amaurosis fugax, Derived From Logistic Regression a Variable B SE Wald Statistic P Odds Ratio 95% CI for Odds Ratio Stenosis PPCS < SGLDM IMC Constant < Abbreviations: PPCS, percentage of pixels of each of the 5 contours of the 0 to 51 gray level spectrum; SGLDM IMC-1, spatial gray level dependence matrices information measure of correlation-1. a Based on the Wald statistic, PPCS1 (a measure of echolucency) was the best parameter. in diagnosing AmF, TIA, and stroke was 100% and 56%, 77%, and 65%, and 71% and 65%, respectively. The distribution of juxtaluminal echolucent plaque components in symptomatic and asymptomatic patients is shown in Table 2; these were more frequent in symptomatic than asymptomatic patients (P <.001). Sensitivity and specificity of a juxtaluminal echolucent plaque component in diagnosing AmF, TIA, and stroke was 93% and 55%,74% and 65%, and 78% and 67%, respectively. Computer-Assisted Plaque Analysis Results of principal component analysis (rotated component matrix) in patients with AmF, TIA, and stroke are summarized in Appendix II. a. Asymptomatic plaques versus plaques associated with AmF (Appendix II and III) Principal components analysis extracted 7 factors (out of 51): percentage of pixels with gray level between 0 and 10 (PPCS1, measure of echolucency), spatial gray level dependence matrices (SGLDM) difference variance, SGLDM correlation, SGLDM information measure of correlation (IMC)-1, percentage of pixels of contour 9 of the gray level spectrum (PPC9), percentage of pixels with gray level between 11 and 20 (PPCS2), and run length run percentage (RP; measures of texture). Because stenosis was more severe in cases with AmF, we performed logarithmic transformation of this variable before entering it in the logistic model. Figure 1. a, Figure comparing the logistic regression output (predicted probability) with the percentage of pixels with gray levels 0 to 10 (PPCS1), using the receiver operating characteristic (ROC) curves method, in distinguishing plaques with amaurosis fugax from asymptomatic ones. The area under the curve was increased from 0.88 (PPCS1) to 0.92 (logistic regression output, see relevant section in Results); the difference was not statistically significant (P ¼.34). b, Scatterplot of PPCS1 versus spatial gray level dependence matrices (SGLDM) information measure of correlation-1 in asymptomatic plaques and those with amaurosis fugax. encountered more frequently in patients with AmF, TIA, and stroke (86%) than patients with no symptoms (47%, P <.001). Sensitivity and specificity of an echolucent plaque Logistic regression model. Several forward and backward models were tested using SGLDM correlation and SGLDM IMC-1. Based on R 2, the best-fit model included stenosis, PPCS1 (measure of echolucency) and SGLDM IMC-1 (measure of texture), as shown in Table 3. The remaining variables were rejected. Based on the Wald statistic, the echolucency measure PPCS1 was more important than texture. Hosmer and Lemeshow test was not significant (P ¼.63), indicating that the model fits adequately. Correct classification rate was 88%. The diagnostic validity of the logistic regression output (predicted probability) was further tested with the ROC curves method and compared with PPCS1. The area under the curve (95% CI) of the regression output was 0.92 ( ), (P <.001)andat sensitivity 77%, specificity was 92%. Logistic regression

5 Kakkos et al 321 Table 4. Independent Parameters Associated With Transient Ischemic Attacks, Derived From Logistic Regression a Variable B SE Wald Statistic P Odds Ratio 95% CI for Odds Ratio PPCS Skewness SGLDM correlation Constant Abbreviations: PPCS, percentage of pixels of each of the 5 contours of the 0 to 51 gray level spectrum; SGLDM, spatial gray level dependence matrices. a The Wald statistic indicates that SGLDM correlation (a measure of texture) is more important than the other parameters. improved the diagnostic value, but the overall difference was small and the results not significant (P ¼.34, Figure 1a); this indicates that echolucency is the main parameter that differentiates the 2 groups. A scatterplot of PPCS1 and SGLDM IMC-1 is shown in Figure 1b. b. Asymptomatic plaques versus plaques associated with TIA (Appendix II and IV) Principal components analysis extracted the following variables: PPCS1, SGLDM difference variance (logarithmic form), SGLDM correlation, PPCS2, run length RP, PPC9, and skewness (logarithmic form). Logistic regression model. Forward and backward models using the Wald method were constructed and led to the same results. The model included PPCS1 (measure of echolucency), skewness, and SGLDM correlation (both measures of texture), as shown in Table 4. Based on the Wald statistic, textural features were more important than echolucency. Hosmer and Lemeshow test was not significant (P ¼.36). Correct classification rate was 77%. The diagnostic validity of the logistic regression output (predicted probability) was further tested with the ROC curves method, and compared with PPCS1, SGLDM correlation, and skewness alone. Logistic regression output demonstrated an increased area under the curve (0.82, 95% CI , P <.001) and improved diagnostic value when compared with PPCS1 (0.72, P <.001), SGLDM correlation (0.70, P <.001), and skewness (0.64, P ¼.014) alone, as shown in Figure 2a, while at sensitivity of 75%, specificity was 74%; the difference between logistic regression output and these 3 parameters was statistically significant (P ¼.009, P ¼.04, and P <.001, respectively). A scatterplot of skewness and SGLDM correlation is shown in Figure 2b. Figure 2. a, Figure comparing the logistic regression output (predicted probability) with percentage of pixels with gray levels 0 to 10 (PPCS1), spatial gray level dependence matrices (SGLDM) correlation and skewness, using the receiver operating characteristic (ROC) curves method in distinguishing plaques in patients with transient ischemic attack from asymptomatic ones. The area under the curve was increased from 0.72 (PPCS1) to 0.82 (logistic regression output, see relevant section in Results); the difference was statistically significant (P ¼.009). b, Scatterplot of skewness vs SGLDM correlation in asymptomatic plaques and those associated with transient ischemic attacks, the latter being more homogeneous. c. Asymptomatic plaques versus plaques associated with stroke (Appendix II and V) Principal components analysis extracted the following variable features: PPCS1, SGLDM difference variance (logarithmic form), SGLDM correlation, PPC9, PPCS2, and run length RP. Logistic regression model. The model (derived by both forward and backward method) included PPCS1 and PPCS2 (measures

6 322 Angiology 62(4) Table 5. Independent Parameters Associated With Stroke, Derived From Logistic Regression a Variable B SE Wald Statistic P Odds Ratio 95% CI for Odds Ratio PPCS < PPCS SGLDM correlation < Abbreviations: PPCS, percentage of pixels of each of the 5 contours of the 0 to 51 gray level spectrum; SGLDM, spatial gray level dependence matrices. a The Wald statistic indicates that PPCS1 (a measure of echolucency) is more important than the other parameters. of echolucency), and SGLDM correlation (measure of texture; Table 5). Based on the Wald statistic, echolucency was more important than textural features. Hosmer and Lemeshow test was not significant (P ¼.53). Correct classification rate was 79%. The diagnostic validity of the logistic regression output (predicted probability) was further tested with the ROC curves method and compared with PPCS1 and GSM. Logistic regression output demonstrated an increased area under the curve (0.85, 95% CI , P <.001) and improved diagnostic value when compared with PPCS1 (0.79, P <.001),SGLDM correlation (0.72, P <.001), and also with GSM (the conventional measure of echolucency; 0.77, P <.001) alone, while at sensitivity of 75%, specificity was 82%. The difference between the logistic regression model output and the PPCS1 was marginally significant (P ¼.08), while SGLDM correlation and GSM were significantly inferior to the logistic regression model (P ¼.012 and P ¼.018, respectively; Figure 3a). A scatterplot of PPCS1 and SGLDM correlation is shown in Figure 3b. An example of a symptomatic and an asymptomatic carotid plaque outline is shown in Figure 3c. The asymptomatic plaque appears more echogenic (low PPCS1) and heterogeneous (high SGLDM correlation) than the symptomatic. The distribution of predicted probability, derived from the logistic regression model and expressed in deciles, in plaques associated with stroke and asymptomatic patients, is depicted in Figure 4. A comparison of texture/histogram features (including GSM, the conventional measure of echolucency) in symptomatic and asymptomatic patients is shown in Table 6. Symptomatic plaques are more echolucent (high PPCS1) and homogeneous (low SGLDM correlation). Additionally, highly echogenic components (PPC9, probably calcium) were observed more often in asymptomatic plaques. Discussion We successfully identified textural features of ultrasonic images of carotid plaques independently associated with particular symptoms, AmF, TIA, and stroke and improved results achieved with objective simple measures of plaque echodensity, that is, PPCS1. Visual plaque type was also recorded, and its association with patient symptoms confirmed previous findings. 11,13 However, because qualitative or semiquantitative (visual and therefore subjective) methods 21,22 are known to be highly subjective, 23 and vary if the computer or scanner monitor brightness are changed, and therefore of limited diagnostic value, we used only objective features. The good results achieved with the use of the latter (as indicated by the area under the ROC curve being 0.92, 0.82, and 0.85 for patients with AmF, TIA, and stroke, respectively) imply that this method could be used in cases of atypical patient symptoms; detection of an unstable morphology could indicate that symptoms are probably embolic in nature and dictate, therefore, a more aggressive management. In this study, there was no significant difference in the degree of stenosis between symptomatic and asymptomatic patients. We wanted to assess the effect of texture and minimize a possible effect of stenosis severity, on symptomatology. Only patients with AmF had worse stenosis, but this was adjusted in logistic regression analysis. To further reduce variability, all images were captured from the same media (ie videotapes) and standardization procedures were applied to the digitized images. In the current study, median image resolution was similar between symptomatic and asymptomatic cases (17.2 pixels/mm vs 18 pixels/mm, respectively, P ¼.47). Following experiments performed by our group, all images used in this study were resized to 20 pixels/mm, using bilinear interpolation, which showed that this method outperformed less sophisticated methods. 11 Similar results were achieved with a resolution of 30 pixels/mm; additionally, when plaques derived from resized images at 10 or 30 pixels/mm were tested, using neural networks, classification rate fell dramatically (data not shown). This reiterates the importance of resolution standardization. To reduce the large number of texture variables, by eliminating the redundant ones, we used principal components analysis, which is a multivariate method; among others, this process identified PPCS1, a measure of echolucent plaque components (reflecting endoluminal thrombus and/or plaque lipid content or hemorrhage), and PPC9, a measure of echogenic components (having gray level above 234, perhaps due to excess calcium) rather than GSM. GSM expresses overall echodensity, and it is reasonable that its individual constituents predominate in this kind of multivariate analysis. Our study is unique from the point that we have separated different symptoms using texture analysis. We have shown that the combined use of selected textural features was better than features of plaque echolucency alone in distinguishing asymptomatic plaques from those associated with hemispheric symptoms. Because different features were important in the three patients groups, a differential plaque structure can be postulated. Echolucency was the major determinant in patients with AmF or stroke, while plaque homogeneity was more

7 Kakkos et al 323 Figure 4. Figure demonstrating the distribution of predicted probability (expressed in deciles, X axis), derived from the regression model, in plaques associated with stroke or those having no symptoms. Y axis represents plaque count. Figure 3. a, Figure comparing the logistic regression output (predicted probability) with percentage of pixels with gray levels 0 to 10 (PPCS1) and gray-scale median (GSM), using the ROC curves method, in distinguishing plaques associated with stroke from asymptomatic ones. The logistic regression output (see relevant section in Results) was marginally better than PPCS1 as far as the area under the curve was concerned (0.85 versus 0.79, P ¼.08), but significantly better than GSM (0.85 vs 0.77, P ¼.018). b, Scatterplot of PPCS1 vs spatial gray level dependence matrices (SGLDM) correlation in asymptomatic plaques and plaques associated with stroke. c, Figure demonstrating an example of a symptomatic (top) and an asymptomatic carotid plaque (bottom) outline. PPCS1 and SGLDM correlation values are shown next to each plaque. The asymptomatic plaque is more echogenic and heterogeneous than the symptomatic one based on its lower PPCS1 and higher SGLDM correlation, results consistent with visual assessment. important in plaques associated with TIAs. The finding of symptomatic plaques being homogeneous is in agreement with previous studies published by our group. 1,2,21 Plaque homogeneity, using a semiquantitative methodology, has previously been shown to be an independent predictor on logistic regression. 22 In contrast, texture analysis methods used in the current study were all quantitative. To the best of our knowledge, only Elatrozy et al 1 suggested that carotid plaque textural features are independently associated with patient symptoms and demonstrated only a trend toward significance for entropy; possible explanations include the small number of patients or use of a nonstandardized methodology (resolution standardization was performed in the current study). Some previous studies have demonstrated that plaque heterogeneity and not homogeneity was associated with patient symptoms. 24,25 Careful interpretation of these studies and inspection of the ultrasonic images shown in their figures demonstrate that only 2 plaque categories were used: heterogeneous and homogeneous; these probably correspond to echolucent and echogenic plaques. Echolucency as estimated by PPCS1 in plaques associated with stroke indicates an increased content of echolucent components, and it could indicate the presence of a more virulent plaque structure, responsible for the severity of symptoms. A large histological study showed that plaques removed 60 days after the most recent event were more unstable after a stroke than after a TIA. 26 However, the instability persisted after a TIA, and plaques removed >180 days after the most recent event were less unstable after a stroke than after a TIA. 26 Information measure of correlation-1 was an independent predictor of AmF in the current study. This feature has been shown previously by our group to be associated with ipsilateral brain infarction in patients with hemispheric symptoms; 14 this study analyzed 54 plaques, 35 of which are included in the current study. Although we previously found that texture analysis can

8 324 Angiology 62(4) Table 6. Results of Texture/Histogram Analysis in Symptomatic and Asymptomatic Patients, Shown As Median (Interquartile Range) a PPCS1 PPCS2 PPC9 SGLDM Correlation SGLDM Difference Variance Runlength RP GSM No symptoms 19.5 ( ) 15.9 ( ) ( ) ( ) 11.6 ( ) 19.5 ( ) 28.1 ( ) AmF 66.0 ( ) b 13.5 ( ) ( ) b ( ) b 9.0 ( ) 14.5 ( ) c 3.3 ( ) b TIA 40.7 ( ) b 14.6 ( ) ( ) ( ) b 12.7 ( ) 17.5 ( ) 16.7 ( ) b Stroke 49.2 ( ) b 15.3 ( ) ( ) b ( ) b 9.1 ( ) c 16.9 ( ) 10.9 ( ) b P value < < <.001 Abbreviations: PPCS, percentage of pixels of each of the 5 contours of the 0 to 51 gray level spectrum; PPC, percentage of pixels of each of the 10 contours of the 0 to 255 gray level spectrum; SGLDM, spatial gray level dependence matrices; RP, run percentage; GSM, gray scale median; AmF, amaurosis fugax; TIA, transient ischemic attack. a GSM, the conventional measure of echolucency is also included. Statistical significance was assessed with the Kruskal-Wallis test. b <.01 using Mann-Whitney test in comparison with asymptomatic plaques. c <.05 using Mann-Whitney test in comparison with asymptomatic plaques. identify symptomatic plaques (TIA and stroke, analyzed together), it was not known whether texture analysis can identify plaques associated with different hemispheric symptomatology, analyzed separately. The previous data set was too small to answer this interesting question, and it did not include plaques associated with AmF and asymptomatic plaques. Also, some strokes/tias might be caused by small emboli not producing a large infarct. The current study extends this previous work with a much larger data set (188 plaques) of asymptomatic plaques and plaques with different hemispheric symptoms analyzed separately. The end-points of the present study (different symptomatology) and the previous study 14 (CT brain infarcts) are different. Although the biologic behavior of carotid plaques associated with AmF and ipsilateral brain infarction is different, from the sonographic point of view these plaque categories are both known to be highly echolucent. 2,14 The observation that PPCS2 (gray level between 11 and 20, diagnostic of lipid core) 27 was associated with plaques producing stroke could be indicative of complex atheroma formation. It is apparent from the 3 scatterplots shown in Figures 1b, 2b, and 3b that symptomatic plaques are scattered widely, while asymptomatic plaques are confined to a small area. This is in contradiction to the present belief that asymptomatic plaque morphology varies from unstable echolucent plaques to stable echogenic plaques. A possible explanation is a change in plaque morphology coinciding with the acute event, due to plaque rupture, thrombosis, and/or hemorrhage. The obvious overlap could be explained by the fact that some asymptomatic plaques have ultrasonic features indicative of vulnerability and might become symptomatic in the near future (all symptomatic plaques were asymptomatic before they became symptomatic); this should be addressed by prospective natural history studies. The seemingly complex association of the various independent factors prompts the use of artificial neural networks. These computer programs are trained with actual cases and used to predict the group or future behavior of new cases. Neural networks are particularly useful in pattern recognition using textural analysis, 28 especially in nonlinear situations, 29 like those we found. The use of neural networks was indeed successful in additional experiments we performed, and the results were comparable with those achieved by the otherwise tedious logistic regression. The overall diagnostic yield of 83% for the 3 symptom groups was better than 75% that was previously found with less sophisticated techniques of texture analysis. 5 This could also be possible because patients were separated into groups according to symptoms and could be improved further by using more efficient software. The current study extends previous work on imaging of atherosclerosis, and it may stimulate further research using the objective and accurate method of image analysis to identify early high-risk as opposed to low-risk plaques (regardless of the degree of stenosis). Furthermore, image analysis could be used to monitor plaque remodeling after different events. Histological studies have showed that plaques evolve differently after different events. 26,34 Plaque inflammation and overall instability persist with time after a TIA but decrease with time after a stroke, suggesting that the nature of the underlying pathology may differ. 26,34 Furthermore, our group showed for the first time that metallothioneins (antioxidant proteins expressed in response to injury) are overexpressed in unstable/echolucent plaques with advanced histological lesions. 35 These differences are probably reflected in the different textural features in plaques associated with different symptomatology. In this context, image analysis could facilitate more appropriate management of the patients, with those at low-risk spare an unnecessary, costly, and potentially dangerous carotid intervention (endarterectomy or angioplasty). Furthermore, image analysis could facilitate monitoring of the efficacy of medical treatment with cardioprotective medications. Longitudinal studies are needed to address these issues. Inherent limitations of our study include its cross-sectional nature; the value of the significant features should be verified by prospective natural history studies of symptomatic patients not operated on. Due to the small number of patients in each group, we were not able to perform a validation analysis (usually performed with neural networks). Also, due to the small sample in each group and the similar time that the ultrasound study was performed in each group after an event in our study, an analysis looking at change in textural features

9 Kakkos et al 325 over time after the event was not possible. Additional limitations of our study are related to the fact that 2D ultrasound images were generated and tested. Atherosclerotic plaques are 3D structures and it is possible that important parts of plaque contents may have been missed because they were out of plane during ultrasonography. Plaque texture analysis from 3D ultrasound could be tested further. 36 The percentage of the plaque that was excluded and not evaluated due to calcification has not been recorded. A screening log was not kept at the time of recruitment; however, all consecutive eligible patients who consented were included in the study. Another limitation is that there were missing data in terms of demographic characteristics and medication use, and therefore we were not able to adjust for those in the statistical analysis. However, Rothwell et al suggested that the predisposition to irregularity and rupture (instability) of atherosclerotic plaques are independent of established, traditional vascular risk factors (but rather local and systemic factors of infection and inflammation). 37 Furthermore, the recent largest ever histological study of carotid endarterectomy specimens by Peeters et al. 38 showed that in both patients with stroke or TIA, macrophages changed over time in a very similar manner in those on and those not on statins, suggesting that plaque remodeling over time was not affected by treatment with statins. Also, although aortic arch atheroma could account for some symptomatic cases (even in patients with severe carotid stenosis), patients were not routinely screened for that. In conclusion, specific texture features of carotid plaque ultrasonic images describing echolucency and texture are associated with the presence of ipsilateral hemispheric symptoms. This amplifies the use of ultrasound in the detection of unstable carotid plaques. Natural history studies having the occurrence of symptoms as an endpoint are needed to confirm the prognostic value of texture analysis. Appendix I Histogram Measures 14 Total number of pixels, percentage of pixels below gray level 30 (PP < 30), percentage of pixels below gray level 50 (PP < 50), percentage of pixels of each of the 10 contours of the 0 to 255 gray level spectrum (PPC1-PPC10; PPC1:0-25, PPC2:26-51, etc), the first 2 contours (gray level 0-51) analyzed further into 5 subcontours (PPCS1-PPCS5; PPCS1: 0-10, PPCS2: 11-20, etc). First-Order Gray Level Parameters 15,16 Mean gray level, variance, median (GSM), mode (expresses the most frequent value), kurtosis, skewness, energy, and entropy. The Spatial Gray Level Dependence Matrices (SGLDM) 10 Angular second moment (ASM), contrast, correlation, variance (sum of squares), inverse difference moment (IDM) homogeneity, sum average, sum variance, sum entropy, entropy, difference variance, difference entropy, information measure of correlation-1 (IMC-1) and information measure of correlation-2 (IMC-2). Angular second moment is a measure of homogeneity of the image. Homogeneous images have very few dominant gray-tone transitions, which result into higher readings. Contrast is a measure of the contrast or the amount of local variations present in an image. Images with large neighboring gray level differences are associated with high contrast. Correlation is a measure of graytone linear-dependencies in the image and heterogeneity. Heterogeneous images have higher correlation values. Gray Level Difference Statistics (GLDS) 17 Entropy, contrast, mean, ASM homogeneity, energy. Gray Level Run Length Statistics 18 Short run emphasis (SRE), gray level distribution (GLD), run length distribution (RLD), long run emphasis (LRE), run percentage (RP). The Fourier Power Spectrum (FPS) 17 Radial and angular features.

10 326 Angiology 62(4) Appendix II Results of Principal Component Analysis (Rotated Component Matrix) in Patients With Amaurosis Fugax, TIA, and Stroke a Amaurosis Fugax TIA Stroke Feature Loading Factor Feature Loading Factor Feature Loading Factor Factor 1 PPCS1 b.92 PPCS1 c.93 PPCS1 d.88 Factor 2 SGLDM difference.95 SGLDM difference.94 SGLDM difference.94 variance (logarithmic form) variance (logarithmic form) variance (logarithmic form) Factor 3 SGLDM correlation.79 SGLDM correlation.86 SGLDM correlation.84 Factor 4 PPC9.93 PPCS2.75 PPC9.94 Factor 5 PPCS2.78 Run length RP.91 PPCS2.83 Factor 6 Run length RP.89 PPC9.93 run length RP.92 Factor 7 PPC2, PPCS3 and PPCS4 <.6 PPC2 and PPCS4 <.6 PPC2 and PPCS4 <.6 Factor 8 - Skewness (logarithmic form).66 - Abbreviations: TIA, transient ischemic attack; PPCS, percentage of pixels with gray; SGLDM, spatial gray level dependence matrices. a Each analysis included also the asymptomatic patients. Total variability explained: b 40%. c 39%. d 35%. Appendix III Results of variable coding in patients with amaurosis fugax and asymptomatic patients (for abbreviations see Appendix I) Percentage of pixels with gray levels 0 to 10 (PPCS1): Amaurosis fugax prevalence in quartile 1 was 5% (1 of 20), in quartiles 2 to 3 24% (10 of 41) and in quartile 495% (19 of 20). Variable was recoded using the corresponding amaurosis fugax prevalence and treated as numerical. Spatial gray level dependence matrices (SGLDM) difference variance: Amaurosis fugax prevalence in quartile 1 was 55% (11 of 20), in quartiles 2 to 3, 34% (14 of 41), and in quartile 4, 25% (5 of 20). Variable was recoded using the corresponding amaurosis fugax prevalence and treated as numerical. Spatial gray level dependence matrices correlation: Amaurosis fugax prevalence in quartile 1 was 85% (18 of 21) vs 20% (12 of 60) in quartiles 2 to 4. Variable was recoded and treated as categorical. Spatial gray level dependence matrices information measure of correlation (IMC)-1: Amaurosis fugax prevalence in quartile 4 was 52% (11 of 21) vs 32% (19 of 60) in quartiles 1 to 3. Variable was recoded and treated as categorical. Percentage of pixels of contour 9 of the gray level spectrum (PPC9): Amaurosis fugax prevalence in deciles 1 to 7 was 48.3% (28 of 58) vs 8.7% (2 of 23) in deciles 8 to 10. Variable was recoded and treated as categorical. Percentage of pixels with gray levels 11 to 20 (PPCS2): Amaurosis fugax prevalence in quartiles 1 and 4 was 25% (10 of 40), in quartile 2 62% (13 of 21), and in quartile 3 35% (7 of 20). Variable was recoded using the corresponding amaurosis fugax prevalence and treated as numerical. Run length RP: Amaurosis fugax prevalence in quartiles 1 and 2 was 51% (21 of 41), in quartile 3 30% (6 of 20), and in quartile 4 15% (3 of 20). Variable was recoded using the corresponding amaurosis fugax prevalence and treated as numerical. Carotid artery stenosis: Ten stenosis groups (deciles) were identified; amaurosis fugax rate was 56% (15 of 27) in 94% stenosis, 42% (13 of 31) in stenosis 80% to 90%, and 9% (2/23) in stenosis groups <80% and >94%. Variable was recoded using the corresponding amaurosis fugax prevalence and treated as numerical. Appendix IV Results of variable coding in patients with transient ischemic attacks (TIAs) and asymptomatic patients. Percentage of pixels with gray levels 0 to 10 (PPCS1): TIA prevalence showed a linear trend across the variable deciles (increasing from 30% to 90%). Variable was recoded using the corresponding TIA prevalence and treated as numerical. Spatial gray level dependence matrices (SGLDM) difference variance, percentage of pixels of contour 9 of the

11 Kakkos et al 327 gray level spectrum (PPC9), percentage of pixels with gray levels 11 to 20 (PPCS2), and run length RP: No difference in TIA rate across the variable deciles or quartiles. Variables were not recoded and treated as numerical. Spatial gray level dependence matrices correlation: TIA prevalence in quartile 1 was 85% (22 of 26) vs 39% (30 of 77) in quartiles 2 to 4. Variable was recoded and treated as categorical. Skewness: TIA prevalence in quartiles 1 to 3 was 42% (32 of 77) vs 77% (20 of 26) in quartile 4. Variable was recoded and treated as categorical. Appendix V Results variable coding in patients with stroke and asymptomatic patients. Percentage of pixels with gray levels 0 to 10 (PPCS1): Stroke prevalence showed a linear trend across the variable deciles (increasing from 0% to 100%). The variable was recoded using the corresponding decile number and treated as numerical. Spatial gray level dependence matrices (SGLDM) difference variance: Stroke prevalence in decile 1 was 90% (9 of 10), 68% (15 of 22) in deciles 2 to 3 and 42% (31 of 74) in deciles 4 to 10. The variable was recoded (3, 2, and 1, respectively) and treated as numerical. Spatial gray level dependence matrices correlation: Stroke prevalence showed a linear trend across the variable deciles (decreasing from 90% to 30%). The variable was recoded using the corresponding decile number and treated as numerical. Percentage of pixels of contour 9 of the gray level spectrum (PPC9): Stroke prevalence in decile 10 was 0% (0 of 10) and in deciles 1 to 9 this was 57% (55 of 96). The variable was recoded and treated as categorical. Percentage of pixels with gray levels 11 to 20 (PPCS2): Stroke prevalence was 36% (15 of 42) in deciles 1 to 2 and 9 to 10, accumulated, while in deciles 3 to 8 this was 63% (40 of 64). The variable was recoded using the corresponding stroke prevalence and treated as categorical. Run length RP: Stroke prevalence in decile 10 was 80% (8 of 10) and in the remaining deciles this was 49% (47 of 96). The variable was recoded and treated as categorical. Declaration of Conflicting Interests The author(s) declared no conflicts of interest with respect to the authorship and/or publication of this article. Funding The author(s) received no financial support for the research and/or authorship of this article. References 1. Elatrozy T, Nicolaides A, Tegos T, Griffin M. The objective characterisation of ultrasonic carotid plaque features. Eur J Vasc Endovasc Surg. 1998;16(3): Tegos TJ, Mavrophoros D, Sabetai MM, et al. Types of neurovascular symptoms and carotid plaque ultrasonic textural characteristics. JUltrasoundMed. 2001;20(2): Mazzone AM, Urbani MP, Picano E, et al. In vivo ultrasonic parametric imaging of carotid atherosclerotic plaque by videodensitometric technique. Angiology. 1995;46(8): El Atrozy T. Preoperative characterisation of carotid plaques. London: University of London; Christodoulou CI, Pattichis CS, Pantziaris M, Nicolaides A. Texture-based classification of atherosclerotic carotid plaques. IEEE Trans Med Imaging. 2003;22(7): Nicolaides AN, Shifrin EG, Bradbury A, et al. Angiographic and duplex grading of internal carotid stenosis: can we overcome the confusion? J Endovasc Surg. 1996;3(2): ECST group. Randomised trial of endarterectomy for recently symptomatic carotid stenosis: final results of the MRC European Carotid Surgery Trial (ECST). Lancet. 1998;351(9113): Rothwell PM, Villagra R, Gibson R, Donders RCJM, Warlow CP. Evidence of a chronic systemic cause of instability of atherosclerotic plaques. Lancet. 2000;355(9197): Elatrozy T, Nicolaides A, Tegos T, Zarka AZ, Griffin M, Sabetai M. The effect of B-mode ultrasonic image standardisation on the echodensity of symptomatic and asymptomatic carotid bifurcation plaques. Int Angiol. 1998;17(3): Haralick R, Shanmugam K, Dinstein I. Textural features for image classification. IEEE Trans Syst Man Cybern. 1973;SMC 3: Kakkos SK, Nicolaides AN, Kyriacou E, Pattichis CS, Geroulakos G. Effect of zooming on texture features of ultrasonic images. Cardiovasc Ultrasound. 2006;4: Geroulakos G, Ramaswami G, Nicolaides A, et al. Characterization of symptomatic and asymptomatic carotid plaques using highresolution real-time ultrasonography. Br J Surg. 1993;80(10): Pedro LM, Pedro MM, Goncalves I, et al. Computer-assisted carotid plaque analysis: characteristics of plaques associated with cerebrovascular symptoms and cerebral infarction. Eur J Vasc Endovasc Surg. 2000;19(2): Kakkos SK, Stevens JM, Nicolaides AN, et al. Texture analysis of ultrasonic images of symptomatic carotid plaques can identify those plaques associated with ipsilateral embolic brain infarction. Eur J Vasc Endovasc Surg. 2007;33(4): Kadah Y, Farag AA, Zurada JM, Badawi AM, Youssef ABM. Classification algorithms for quantitative tissue characterization of diffuse liver disease from ultrasound images. IEEE Trans Med Imag. 1996;15(4): Wilhjelm JE, Gronholdt ML, Wiebe B, Jespersen SK, Hansen LK, Sillesen H. Quantitative analysis of ultrasound B-mode images of carotid atherosclerotic plaque: correlation with visual classification and histological examination. IEEE Trans Med Imaging. 1998;17(6): Wenska J, Dryer CR, Rosenfeld A. A comparative study of texture measures for terrain classification. IEEE Trans Syst Man Cybern. 1976;SMC 6:

12 328 Angiology 62(4) 18. Galloway M. Texture classification using gray level run lengths. Comp Graph Image Process. 1975;4(2): Metz CE. Basic principles of ROC analysis. Semin Nucl Med. 1978;8(4): Hanley JA, McNeil BJ. A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology. 1983;148(3): el-barghouti N, Nicolaides AN, Tegos T, Geroulakos G. The relative effect of carotid plaque heterogeneity and echogenicity on ipsilateral cerebral infarction and symptoms of cerebrovascular disease. Int Angiol. 1996;15(4): Tegos TJ, Stavropoulos P, Sabetai MM, Khodabakhsh P, Sassano A, Nicolaides AN. Determinants of carotid plaque instability: echoicity versus heterogeneity. Eur J Vasc Endovasc Surg. 2001;22(1): Nicolaides AN, Kakkos SK, Griffin M, et al. Effect of image normalization on carotid plaque classification and the risk of ipsilateral hemispheric ischemic events: results from the asymptomatic carotid stenosis and risk of stroke study. Vascular. 2005;13(4): AbuRahma AF, Kyer PD 3rd, Robinson PA, Hannay RS. The correlation of ultrasonic carotid plaque morphology and carotid plaque hemorrhage: clinical implications. Surgery. 1998;124(4): AbuRahma AF, Wulu JT Jr, Crotty B. Carotid plaque ultrasonic heterogeneity and severity of stenosis. Stroke. 2002;33(7): Redgrave JN, Lovett JK, Gallagher PJ, Rothwell PM. Histological assessment of 526 symptomatic carotid plaques in relation to the nature and timing of ischemic symptoms: the Oxford plaque study. Circulation. 2006;113(19): Madycki G, Staszkiewicz W, Gabrusiewicz A. Carotid plaque texture analysis can predict the incidence of silent brain infarcts among patients undergoing carotid endarterectomy. Eur J Vasc Endovasc Surg. 2006;31(4): Mayorga M, Ludeman LC. Neural Nets for Determination of Texture and Its Orientation. ICASSP International Conference on Acoustics, Speech, and Signal Processing, Toronto, Ontario, Canada; IEEE : Sujana H, Swarnamani S, Suresh S. Application of artificial neural networks for the classification of liver lesions by image texture parameters. Ultrasound Med Biol. 1996;22(9): Golemati S, Tegos TJ, Sassano A, Nikita KS, Nicolaides A. Echogenicity of B-mode sonographic images of the carotid artery: work in progress. J Ultrasound Med. 2004;23(5): Hara T, Yamada S, Hayashi T, et al. Accuracy of nonstenotic coronary atherosclerosis assessment by multi-detector computed tomography. Circ J. 2007;71(6): Motoyama S, Kondo T, Anno H, et al. Atherosclerotic plaque characterization by 0.5-mm-slice multislice computed tomographic imaging. Circ J. 2007;71(3): Seo Y, Watanabe S, Ishizu T, et al. Echolucent carotid plaques as a feature in patients with acute coronary syndrome. Circ J. 2006; 70(12): Peeters W, Hellings WE, de Kleijn DP, et al. Carotid atherosclerotic plaques stabilize after stroke: insights into the natural process of atherosclerotic plaque stabilization. Arterioscler Thromb Vasc Biol. 2009;29(1): Daskalopoulou SS, Daskalopoulos ME, Theocharis S, et al. Metallothionein expression in the high-risk carotid atherosclerotic plaque. Curr Med Res Opin. 2007;23(3): Coleman DP, Rakenbrandt F, Pugh ND, Crawford DC, Woodcock JP. Development and validation of an in vivo analysis tool to identify changes in carotid plaque issue types in serial 3-D ultrasound scans. Ultrasound Med Biol. 2005; 31(3): Rothwell PM, Villagra R, Gibson R, Donders RC, Warlow CP. Evidence of a chronic systemic cause of instability of atherosclerotic plaques. Lancet. 2000;355(9197): Peeters W, Hellings WE, de Kleijn DP, et al. Carotid atherosclerotic plaques stabilize after stroke: insights into the natural process of atherosclerotic plaque stabilization. Arterioscler Thromb Vasc Biol. 2009;29(1):