The North American Symptomatic Carotid Endarterectomy Trial (NASCET) and the European Carotid Surgery Trial (ECST) have shown that surgery in symptomatic patients with severe internal carotid artery stenosis results in a six-fold reduction of stroke risk [1, 2]. However, patients who do not have severe stenoses and patients who are asymptomatic can also go on to develop stroke. It is, therefore, important to be able to determine whether any of these patients have carotid plaques which are high-risk or unstable. Ultrasound greyscale median (GSM) is commonly used to quantify the ultrasound appearance of carotid plaques, and several studies have found that it may be valuable for predicting the risk of cerebrovascular events. In particular, statistically significant associations have been reported between plaque GSM and the presence of cerebrovascular symptoms [3, 4], cerebral infarction in symptomatic and asymptomatic patients [5–7], recurrent cerebrovascular events before undergoing carotid endarterectomy [8], and the overall risk of stroke in symptomatic patients [9], asymptomatic patients [10], and during and after carotid artery stenting [11].

GSM measurements currently have poor reproducibility across studies. This can be partly attributed to the differences in the acquisition settings used in separate studies. In order to reduce this variability, investigators have attempted to standardise ultrasonic images of carotid plaques by specifying certain acquisition settings to be used for carotid artery scanning and normalizing the resultant ultrasound images [12]. However, existing studies typically measured GSM on still ultrasound images, and thus ignored any variations that may have been observed on a frame-by-frame basis. The purpose of this study was to establish the existence of and investigate the nature and extent of any frame-by-frame variations in the plaque GSM using a novel technique for tracking plaque boundaries in ultrasound image sequences. We hypothesized that variations in the GSM of carotid artery plaques may occur due to deformation of the plaque during the cardiac cycle, and other confounding factors such as out-of-plane plaque, patient or probe motion. Furthermore, it was hypothesized that plaques of different composition and morphology may exhibit different inter-frame variations in GSM in otherwise equivalent hemodynamic circumstances and hence the measurement of these variations may give useful insight into the dynamic behaviour of plaques and help identify vulnerable plaques.

Plaque outlines could be tracked successfully in a variety of different configurations (Figure 2). Across all plaque samples, the un-normalized plaque GSM, averaged across all frames, ranged between 26 and 112 (mean 47, Table 1). Plaque areas ranged between 7 mm² and 92 mm² (mean 30 mm²). The mean inter-frame coefficient of variation (s.d./mean) of GSM was 5.2% (s.d. 2.5%) while that of plaque area was 4.2% (s.d. 2.9%). In relation to the normalized GSM obtained using the NORM1 method, the corresponding mean GSM figures ranged between 24 and 96 (mean 46). The mean inter-frame coefficient of variation was the same as without normalization (5.2%) but the standard deviation was slightly larger (2.6%). Normalization using the second normalisation technique (NORM2) for plaques px1, px2, px3, px19 (excluding the region of acoustic shadowing) and px22 resulted in larger coefficients of variation (4.8%, 9.7%, 6.7%, 3.8% and 4.9% respectively) compared to the un-normalized and NORM1 normalized coefficients of variation (Table 1).

Periodic variations with frequencies of the order of 60/min in either or both of the plaque GSM or area waveforms were observed for 12 plaques (50%) but not observed for 12 other plaques (50%, Table 1). Three plaques were excluded from this analysis as they had short acquisition sequences.

Overall, 13 of the 27 plaques (48%) exhibited inter-frame variations in GSM of greater than 5% measured as the inter-frame coefficient of variation of GSM in both the un-normalized and NORM1 normalized cases. In contrast, only 6 of the 27 plaques (22%) had inter-frame coefficients of variation in plaque area of greater than 5% (Table 1, Figure 3).

Normalisation using NORM1 did not appear to change the shape of the GSM variation waveform but caused a translation along the y-axis (Figure 4). After normalization, the mean GSM was lower for some plaques, and higher for others (Figure 5a). The coefficients of variation were predominantly the same (Figure 5b), yet for some plaques, NORM1 also changed the coefficient of variation (Figure 5b).

The correlation between the inter-frame coefficients of variation in un-normalized plaque GSM and cross-sectional area (Figure 6) was not statistically significant (Spearman’s rho 0.36, p = 0.07). Testing the influences on the extent of the inter-frame variations seen in the un-normalized GSM, of (a) the mean inter-frame, un-normalized GSM, (b) the mean inter-frame plaque area, and (c) the extent of inter-frame variations seen in plaque area, with the extents taken as the standard deviation of inter-frame values, identified the mean inter-frame GSM as the only statistically significant factor at the 5% significance level (Table 2).

The mean, normalized GSM differed significantly between the symptomatic and asymptomatic groups (p = 0.002) but the parameters based on the inter-frame variations in the normalized GSM did not (p = 0.48 for the inter-frame standard deviation of normalized GSM and p = 0.42 for the coefficient of variation of normalized GSM, Figure 7).

Our plaque GSM and area measurements showed good reproducibility (Table 3) and were broadly comparable to those obtained using manual delineation (Table 4). The mean intra-observer coefficients of variation for the eight selected plaque samples were 1.4% for the measurement of the un-normalized GSM, 2.4% for the plaque area, and 2.8% for the NORM1 normalized GSM (Table 3). Manual delineation results showed a greater amount of variation (mean coefficients of variation were 7.7% for the un-normalized GSM and 8.0% for the plaque area compared with 5.4% and 4.0%, respectively, for our method), due to the greater subjectivity of the manual delineation process (Table 4). However, the mean difference in GSM measurements between the two techniques was 0.1 grey levels (Figure 8) and did not differ significantly from zero (p = 0.77, t-test). The 95% limits of agreement were −7.9 grey levels to +8.0 grey levels.

Plaque sample	GSM	Area	NORM1
px2	1.7% (0.24)	2.9% (0.37)	2.3% (0.29)
px4	0.5% (0.18)	2.3% (0.11)	4.4% (1.39)
px11	1.6% (0.48)	1.4% (0.24)	4.7% (1.53)
px26	1.1% (0.22)	1.5% (0.26)	1.2% (0.30)
px21	1.4% (0.32)	2.5% (0.25)	2.4% (0.56)
px22	1.5% (0.45)	2.9% (0.29)	1.7% (0.43)
px12	1.8% (0.44)	3.3% (0.22)	2.4% (0.66)
px5	1.4% (0.23)	2.3% (0.23)	3.3% (0.46)
Column means	1.4% (0.32)	2.4% (0.24)	2.8% (0.70)

Plaque sample	Our method						Manual delineation
	GSM			Area			GSM			Area
	mean	s.d.	CV	mean (mm2)	s.d. (mm2)	CV	mean	s.d.	CV	mean (mm2)	s.d. (mm2)	CV
px2	32.3	2.80	8.7%	27.9	1.12	4.0%	28.7	3.88	13.5%	27.5	2.10	7.6%
px4	76.9	2.22	2.9%	11.3	1.05	9.3%	78.1	4.57	5.9%	11.0	2.03	18.4%
px11	70.2	1.27	1.8%	38.0	0.23	0.61%	73.0	3.36	4.6%	37.4	0.40	1.1%
px26	43.4	1.49	3.4%	39.5	0.94	2.4%	45.9	2.27	4.9%	36.9	1.61	4.4%
px21	52.4	2.35	4.5%	22.7	1.09	4.8%	51.3	2.86	5.6%	22.1	1.80	8.1%
px22	67.4	2.15	3.2%	21.5	0.80	3.7%	67.6	3.54	5.2%	21.3	1.75	8.2%
px12	54.0	3.72	6.9%	14.9	0.39	2.6%	52.9	4.47	8.5%	15.0	1.23	8.2%
px5	36.9	4.31	11.7%	22.0	0.96	4.4%	36.0	4.66	13.0%	22.4	1.71	7.7%
Column means	54.2	2.54	5.4%	24.7	0.82	4.0%	54.2	3.70	7.7%	24.2	1.58	8.0%

GSM values are un-normalized and CV is the coefficient of variation (s.d./mean).

Our investigation highlighted variations in the GSM and area of plaques when measured on a frame-by-frame basis throughout ultrasound image sequences. Image normalisation did not reduce the extent of the GSM variations and in some cases resulted in greater variation. These results demonstrate that frame-by-frame variations in the plaque GSM cannot be offset by applying normalisation factors based on the selection of blood and adventitia regions in one of the frames. Furthermore, selecting separate reference regions in all images introduced an additional source of variability due to the subjective nature of the process. The reference regions were user-selected and not computerized as they are easily identified by the operator and it would have been difficult to ensure the accuracy of a computerized selection. In NORM1 normalisation, the coefficients of variation for GSM changed after normalization for some plaques. This occurred when the blood reference regions had non-zero GSM, causing an intercept to be introduced into the linear relationship between the normalized and un-normalized greyscale values. Furthermore, as GSM values are limited to the range 0 to 255, normalization could result in the clipping of the GSM values outside this range, thus affecting the coefficients of variation. The increased coefficients of variation for GSM, in the case of the NORM2 normalisation, provided evidence that the frame-by-frame variations seen in GSM were unlikely to be, at least significantly, due to a general temporal variability in the overall image brightness.

The extent of the GSM variations seen in the study were similar in magnitude to those reported by Elatrozy et al. [12], who found that, after normalisation, the coefficient of variation among 4 different observers was 4.7% for the GSM of plaques. However, the variations captured by that study did not include variations that may have been observed to the selection of different still images as each of the four observers appeared to have used the same image to assess the GSM. Therefore, the true inter-observer variabilities may have been greater than that suggested by the results of that study.

The findings of our study have two implications. Firstly, in the case of studies which have already considered intra/inter-observer variabilities, the variabilities found may have been under-estimated unless the same image frames were not used in multiple assessments of the GSM. Secondly and conversely, in the case of inter/intra-sonographer or across-study variabilities of GSM measurements, some of these variabilities may have been due to the selection of different image frames corresponding to the differences in the exact cross-sections being imaged.

The variations seen in the plaque GSM and area may be due to a number of different factors. While changes to the acquisition settings during a single acquisition would not be expected, changes in the plaque GSM could occur, for example, if the distance between the plaque and the transducer face changed during an acquisition. Patient or probe motion may also change the location and orientation of the scan plane with respect to the plaque being imaged, affecting both the measured GSM and the observed cross-sectional area. These are likely to be significant contributors to the variations seen in the GSM and area of plaques in this study. Deformation or compression of the plaque under the pulse pressures may also cause changes in the measured plaque GSM and cross-sectional area and the observation of periodical variations with physiologically reasonable frequencies in the plaque GSM and area for several of the plaques provided evidence to support this hypothesis. However, it should be noted that such cyclic variations could also have been caused by periodic variations in the scan plane location and orientation due to out-of-plane plaque, patient, or probe motion. The poor correlation between the inter-frame coefficients of variation of GSM and plaque area suggested that at least some of the variations seen in the plaque GSM were likely to have occurred due to factors other than changes in the observed plaque area. This was also supported by the results of the regression analysis, which did not highlight the parameters based on the plaque area as being statistically significant contributors to the variabilities seen, across plaque samples, in the extent of the inter-frame GSM variations.

Other factors that may have caused apparent changes in the plaque GSM and cross-sectional area included unclear plaque boundaries (e.g. poor quality image or substantial image noise), which may have caused fluctuations in the detected plaque boundaries. However, the images used in this study were of sufficiently good quality that any variations due to such fluctuations were not thought to be a major contributor to the GSM variations observed.

The statistically significant difference found in the mean GSM of plaques in the symptomatic and asymptomatic groups is in accord with previous findings that have shown symptomatic plaques, in general, had lower GSM [3, 4]. The differences between the two groups in the case of the parameters describing the inter-frame variations in the GSM were not statistically significant which was plausible as out-of-plane plaque, patient, and probe motion appeared to be a significant sources of variation for GSM measurements.

Manual delineation of the plaque boundaries separately for each image frame was found to increase the extent of the frame-by-frame variations observed in the plaque GSM and area (7.7% and 8.0%, respectively, compared with 5.4% and 4.0% for our method) due to the greater subjectivity of the manual delineation process.

The main limitations of our study were the use of two dimensional ultrasound and the absence of any attempts to fix the scan plane location and orientation with respect to the plaque being imaged, other than those measures normally taken in the clinics (e.g. holding the probe fixed and asking the patient to remain still). It should be remembered that the method of ultrasound acquisition commonly used in carotid plaque GSM studies, namely two dimensional ultrasound, provides only a cross section of the whole plaque volume. Since the plaque GSM measured using two dimensional techniques reflects only a cross-section, these measurements are susceptible to variations due to out-of-plane plaque, probe and patient motion. Studies incorporating three dimensional techniques may overcome this limitation and enable further investigation of the nature of any intrinsic frame-by-frame variations in the plaque GSM. Such follow-up studies may also identify whether any inter-frame variations seen in the GSM and volume of plaques can provide additional insight into the dynamic behaviour of carotid plaques, thus improving clinical utility.

Another limitation of our assessment of the plaque GSM and cross-sectional area was with regard to anechoic regions of plaques and regions of plaques in areas of acoustic shadowing. These regions were excluded from analysis. The cross-sectional areas of plaques that had such regions were, thus, under-estimated and neither the plaque area nor the GSM reflected the true values. The excluded regions also had an effect on the magnitude of the frame-by-frame variations that were measured for the affected plaques. In the case of anechoic regions of plaques, the inclusion of the anechoic regions would have reduced the magnitude of the variations observed. However, this would have been only because of the absence of echogenicity in these regions. In the case of the regions of acoustic shadowing, these regions need to be excluded from analysis due to the absence or deterioration of plaque texture information resulting from acoustic shadowing. Although Colour Doppler is useful for subjectively defining the plaque-lumen boundary, it is not suitable for quantifying the plaque area throughout the cardiac cycle, since colour filling of the lumen is dependent on blood flow velocity [17]. Nevertheless, our results demonstrated that variations in the plaque GSM and cross-sectional area were observed, in the visible parts of the plaques that were not in regions of acoustic shadowing. Such variations are important as they could lead to an error in a potential diagnostic test that uses the GSM as the selection criterion, particularly for plaques of intermediate echogenicity where a coefficient of variation of 5% may provide enough bias to move the plaque between the high-risk and low-risk groups. As the plaque GSM is not generally used to inform clinical decision making, these variations do not currently have a clinical impact. Nevertheless, the variations should be appreciated for research studies which increasingly utilize the plaque GSM.

Our results did not find the cross-sectional plaque area to affect the extent of the frame-by-frame variations observed in the plaque GSM significantly. A major source of variation in the inter-frame plaque GSM may in fact be the movement of the plaque cross-section with respect to the scan plane and this may be a bigger problem for smaller plaques. However, our results showed that the observed GSM could vary on a frame-by-frame basis substantially for large plaques as well the minor stenoses.

Since previous studies typically quantified GSM in single frames of ultrasound images, the variations found in this study have been previously neglected. Improved attempts to standardise GSM measurements and reduce variability between centres should account for these findings, for example, by performing GSM measurements at peak systolic/diastolic frames or by carrying out an assessment of the average GSM throughout the cardiac cycle. Techniques such as GSM assessment using multiple cross-sectional views of plaques and plaque texture analysis can also be used to improve diagnostic precision compared to a single cross-sectional assessment. The best option would be to carry out these processes in three dimensions, however, three dimensional ultrasound techniques are still under development and are not widely available. It should be noted that we do not propose the technique we have used in our study as a replacement for other methods but we highlight the variations in plaque GSM and cross-sectional area that may be observed on a frame-by-frame basis using single-view, two dimensional ultrasound.

In conclusion, this investigation found that the GSM of carotid artery plaques can vary when measured on a frame-by-frame basis throughout ultrasound image sequences. These variations affect the reproducibility of studies and have implications for the use of GSM as a predictor of cerebrovascular events. Future studies looking at the GSM of carotid artery plaques may need to take these variations into consideration.