Semi-automatic Carotid Intraplaque Hemorrhage Volume Measurement using Three-Dimensional Carotid MRI

Abstract

Purpose:

Develop a semi-automatic method to measure carotid intraplaque hemorrhage (IPH) volume.

Study type:

Retrospective

Population:

Patients scheduled for carotid endarterectomy and patients with 16–79% asymptomatic carotid stenosis by ultrasound.

Field Strength:

3T

Sequence:

Simultaneous noncontrast angiography and intraplaque hemorrhage (SNAP) MRI

Assessment:

A semi-automated volumetric measurement of IPH using signal intensity thresholding of 3D SNAP volume was implemented. Fourteen carotid endarterectomy patients were enrolled to determine the signal intensity threshold of IPH using histology. Thirty-three patients with 16–79% asymptomatic stenosis were scanned twice within one month to evaluate reproducibility. The normalized SNAP intensity with the highest Youden index for predicting IPH on histology was used for thresholding. Scan-rescan reproducibility of IPH measurement was assessed using intraclass correlation coefficient (ICC) and coefficient of variation (CV).

Statistical tests:

Receiver operating characteristic curve, area under the curve, Cohen’s kappa, intraclass correlation coefficient, coefficient of variance (CV), paired t-test.

Results:

IPH detection by the algorithm had substantial agreement with manual review (kappa: 0.92; 95% CI: 0.83, 1.00) and moderate agreement with histology (kappa: 0.55; 95% CI: 0.34, 0.68). IPH volume measurements by the algorithm were strongly correlated with histology (Spearman’s rho = 0.76, p = 0.002). IPH measurements were also reproducible, with ICCs of 0.86 (95% CI: 0.57, 0.96), 0.77 (95% CI: 0.32, 0.94), and 0.99 (95% CI: 0.93, 1.00) for maximum/mean normalized intensity and IPH volume, respectively. The corresponding CVs were 10.6%, 5.2%, and 11.8%.

Data Conclusion:

IPH volume measurements on SNAP MRI are highly reproducible using semi-automatic measurement.

Introduction

Intraplaque hemorrhage (IPH) is a characteristic feature of high-risk atherosclerotic plaques [1]. Prospective natural history studies in various populations have demonstrated a four- to twelve-fold increased risk for cerebrovascular ischemic events in patients with carotid IPH [2–6]. Serial MRI studies further revealed that IPH is a potent driver of atherosclerotic plaque progression[7,8]. Therefore, measurement of IPH on MRI may provide an important biomarker of stroke risk.

While IPH presence or absence is considered a stable biomarker [8–10], the extent of IPH may change. Studies that quantitatively measured the extent of IPH on MRI suggest that IPH is a dynamic pathological process [8,10–12]. For example, in recently symptomatic patients with bilateral IPH, the ipsilateral plaques showed higher T1 signals than the contralateral plaques [13]. Notably, natural history studies have shown that most ischemic strokes due to carotid atherosclerosis are associated with IPH, whereas the short- to mid-term stroke risk in the absence of IPH is minimal [3,14]. Nonetheless, there is substantial heterogeneity among IPH plaques, particularly in asymptomatic patients [13]. Therefore, quantitative measures of IPH signals (volume and intensity) by MRI may contribute to a more precise assessment of clinical risk beyond the mere presence of IPH. When applied in serial studies, such measurements may serve as biomarkers to help understand risk factors and effective therapies for IPH-induced plaque progression.

Reproducible measurement methods for IPH are required to precisely track small changes in IPH volume over time. However, few studies have reported scan-rescan reproducibility of quantitative IPH measurements, while similar measurements of other plaque components have been well documented [15–17]. Limited sample size with IPH has been a factor hampering IPH quantification measurements [18].

Large-coverage, three-dimensional MRI is being adopted in clinical and population studies of IPH [4,19] with increased time demands on manual quantification [10,13] due to the higher resolution and larger coverage. Manual segmentation of IPH is not only time-consuming but also prone to measurement errors due to the irregular and inconspicuous boundaries of IPH areas. A semi-automated approach to IPH segmentation and quantification may reduce both measurement variability and manual effort.

Inversion-recovery prepared gradient echo MRI is used to detect carotid IPH [20]. Large coverage 3D isotropic inversion-recovery with phase-sensitive reconstruction to detect IPH and luminal stenosis concurrently as simultaneous noncontrast angiography and intraplaque hemorrhage (SNAP) MRI [21] is highly T1 weighted and therefore IPH is depicted as hyperintensity on SNAP and is readily detectable. The improved IPH-to-wall contrast in SNAP as compared to traditional magnitude-only images may also improve IPH quantification. These advantages of SNAP suggest that it has the potential for semi-automated IPH processing.

Therefore, this study sought to develop a semi-automated method for carotid IPH measurement and to evaluate the accuracy and scan-rescan reproducibility of IPH measurements using SNAP MRI.

Materials and Methods

Study Design

This retrospective study was performed in compliance with the Health Insurance Portability and Accountability Act, and the study protocol and procedures were approved by the local institutional review board. Written informed consent was obtained from enrolled subjects. Two groups of subjects were recruited. Group A consisted of 14 patients who were scheduled for carotid endarterectomy (CEA). Histology was available for these subjects and was used to determine the signal intensity threshold for detecting IPH. Group B consisted of 33 patients with 16–79% asymptomatic carotid stenosis by ultrasound. These subjects underwent two SNAP scans to study scan-rescan reproducibility of IPH signals using the semi-automatic method. Consecutive screening and enrollment of subjects occurred between November 2011 and November 2013 for Group A, and between June 2013 and February 2016 for Group B.

MR Imaging

All SNAP scans were performed at 3T (Philips, Best, The Netherland) using an eight-channel carotid coil [22]. In Group A, MRI was performed prior to surgery, including SNAP and a conventional 2D multicontrast protocol [15]. In Group B, two SNAP scans were performed within one month. SNAP imaging parameters were as follows [23]: repetition time/echo time = 10/4.8 msec, flip angle = 11°, inversion time = 500 msec, field-of-view = 160×160×32 mm³, scan time = 5.3 minutes. Spatial resolution was 0.8×0.8×0.8 mm³ acquired and 0.4×0.4×0.4 mm³ interpolated.

Histological Analysis

Processing and analysis of CEA specimens were consistent with previous studies [24–26]. Carotid plaques were removed intact during surgery and gently washed with saline. After fixation in formalin and decalcification in 10% formic acid, each plaque was embedded in paraffin en bloc and sectioned axially every 1 mm in the common carotid or 0.5 mm in the internal carotid, due to more complexity and variability in composition and morphology in the internal carotid artery portion. The sections were stained with Hematoxylin-eosin stain and examined under various magnifications [25]. The presence and boundary of IPH were determined by an experienced reader (MSF with 26 years of experience in vascular pathology) blinded to clinical information and MRI findings.

Signal Intensity Normalization

SNAP utilizes phase-sensitive inversion recovery [21,27]. Three sets of images are generated [23]: 1) heavily T1-weighted images (I1); 2) proton density-weighted reference images (I2); and 3) phase-corrected images (SNAP). The phase-corrected images are suitable for characterizing IPH because of the unique image contrast which shows IPH and carotid lumen as strong positive and negative signals, respectively. Other soft tissues have near-zero signal intensities on the SNAP image. Therefore, SNAP images were normalized voxel-by-voxel to the mean signal intensity of sternocleidomastoid muscle (SCM) on I2, sampled using a 2D region-of-interest (ROI) (Figure 1). SNAP MRI has -in coil sensitivity correction, which mitigates influences of coil sensitivity on signal intensity measurements.

Figure 1.

Volumetric processing of SNAP data for quantitative characterization of IPH. (a) SNAP maximum (white color) and minimum (red color) intensity projections were generated and overlaid, after which a 3D ROI (blue box) was interactively defined by selecting two points in coronal and sagittal views, respectively. Note the penetrating ulcer in the internal carotid artery at the level of stenosis (green arrow). (b) The signal intensity of sternocleidomastoid muscle on I2 (proton density-weighted reference image) was sampled for signal intensity normalization (yellow ellipse). (c) SNAP voxels that exceeded the mean signal intensity of sternocleidomastoid muscle on I2 (pre-determined threshold for IPH detection) and were part of any 2×2×2 IPH-positive cubic were classified as IPH-positive voxels (red contours). The two axial images were reformatted at levels of IPH (yellow dashed lines) as shown in (a).

Signal Intensity Threshold for IPH

In Group A, histological sections (10 μm thick slices every 0.5–1 mm) and reformatted axial SNAP images (0.4 mm thick slices every 2 mm) were matched indirectly by referencing to 2D MR images (2 mm thick consecutive slices). Matching between histology and 2D MR images was based on lumen/wall morphology and distance to carotid bifurcation, taking into account possible distortion and shrinkage during histological processing [24]. Matching between SNAP and 2D MR images was solely based on distance to carotid bifurcation. As histological sampling was finer than the spacing between MR slices, multiple histological sections could be matched to a single MR slice. IPH was considered present if IPH was detected in any of the matched histological sections.

The maximum normalized intensity of the artery wall from each SNAP image was compared to IPH presence or absence on matched histological sections to determine the optimal signal intensity threshold for IPH (see Statistical Analysis). For comparison with the semi-automatic measurement described in the next section, the IPH volume of each plaque on histology was calculated as the sum of the areas within the IPH boundary multiplied by the histologic section spacing.

To provide an additional reference standard with which to measure IPH detection performance of the semi-automatic method, the presence/absence of IPH was also determined from the reformatted axial SNAP images by human readers. Three readers (JL with 4 years, JS with 9 years, and a third reader with 8 years of experience in vessel wall MRI) who were blinded to the histological and clinical findings independently reviewed the SNAP images and determined IPH presence by visually inspecting whether there were distinct hyperintense signals within the carotid wall.

Quantitative Characterization of IPH Using Volumetric Image Processing

Based on the signal intensity threshold for IPH determined using histology, a volumetric image processing method was developed for time-efficient, quantitative characterization of IPH (Figure 1). First, maximum and minimum intensity projections were generated and overlaid to facilitate region selection. Second, a 3D box covering the carotid bifurcation was defined interactively on the compound display by selecting two points (left upper and right lower corners) in the coronal and sagittal views, respectively. The 3D ROI enclosed the carotid artery and was wide enough to include any hyperintense areas, if present. Third, the signal intensity of the SCM on I2 was sampled and used to normalize SNAP signal intensities. The signal intensity threshold for IPH was applied to all voxels (0.4×0.4×0.4 mm³) in the 3D ROI to obtain a 3D binary matrix coded with presence or absence of IPH. A single threshold was used for all subjects and all slices within a subject. The 3D binary matrix was then filtered with a 2×2×2 cubic (morphological opening) to exclude isolated voxels (interpolated) given the acquired resolution of 0.8×0.8×0.8 mm³. Maximum normalized intensity was recorded for each artery. IPH volume was calculated as the number of voxels in all IPH regions multiplied by the voxel size. Mean normalized intensity of IPH regions was measured.

Reproducibility Studies

In Group B, subjects were scanned twice within one month. The 3D ROI defined on the first scan was automatically mapped to the reproducibility scan by local rigid registration (MATLAB, R2015a). SCM signal intensity on I2 was measured separately. Maximum normalized intensity, IPH volume, and mean normalized intensity of IPH were then measured automatically.

Statistical Analysis

In Group A, the receiver operating characteristic curve and area under the curve (AUC) were analyzed for an overall measure of diagnostic performance. The normalized SNAP intensity that produced the highest Youden index (sensitivity + specificity - 1) was selected as the optimal threshold for IPH [28]. Sensitivity and specificity for IPH detection of the optimal threshold were estimated at the slice level using leave-one-out cross-validation (leaving out one plaque at each iteration) with the histological assessment as the reference standard. Sensitivity and specificity of the manual MR review for IPH presence were also calculated at the slice level, where IPH was defined to be present if detected by at least two of three independent readers. Overall agreement in IPH detection between semi-automatic MR review, manual MR review, and histological analysis, was evaluated using Cohen’s kappa. Agreement in IPH volume quantitation between SNAP and histology was assessed using Spearman’s correlation coefficient, which is less affected by specimen shrinkage during processing than measurements of absolute agreement like the intraclass correlation coefficient (ICC).

In Group B, Cohen’s kappa was used to assess scan-rescan reproducibility in presence of IPH signals. The ICC and coefficient of variance (CV) were used for evaluating the reproducibility of maximum normalized intensity, IPH volume and mean normalized intensity of IPH. The paired t-test was used to test differences between first and reproducibility scans.

To adjust for repeated measures (multiple slices per artery in Group A; bilateral arteries per subject in Group B), the non-parametric bootstrap and percentile method was used to calculate 95% confidence intervals (CIs). Statistical significance was defined as p<0.05. Statistical analysis was performed using R (version 3.0.2).

Results

Patient Characteristics

Fourteen subjects were recruited for Group A (68.4 years ± 9.7; 12 males) (Table 1). The mean time interval between MRI and surgery was 10.5 days ± 12.0. Thirty-three subjects were recruited for Group B (67.8 years ± 14.7; 23 males) (Table 1). The mean inter-scan time interval was 7.8 days ± 8.9. Sixty-six carotid arteries were thus included in the reproducibility study.

Table 1.

Study Populations

Characteristics	Group A (n = 14)	Group B (n = 33)
Age (y)*	68.4 ± 9.7	67.8 ± 14.7
Male	12 (86)	23 (70)
Symptomatic	3 (21)	0 (0)
Hypertension	11 (79)	27 (82)
Hyperlipidemia	14 (100)	31 (94)
Diabetes	3 (21)	6 (18)
Current smoker	5 (36)	6 (18)
Current statin user	14 (100)	26 (79)

Note: All numbers except age are numbers of patients, with percentages in parentheses.

^*Mean age ± standard deviation

Signal Intensity Threshold for Predicting IPH in Histology

In Group A, a total of 121 SNAP slices were matched with histological sections, of which 49 (40%) slices from 13 (93%) plaques had histology-detected IPH. The overall AUC for IPH detection was 0.78 (95% CI: 0.62, 0.90). The signal intensity threshold based on the Youden index was found to be 100% (99.5% precisely) of SCM signal intensity on I2 (Figure 2). IPH was detected in 34 MR slices from 7 plaques using this threshold.

Figure 2.

Signal intensity threshold optimization using histology. (a) The optimal signal intensity threshold for IPH detection on SNAP (phase-corrected image) was chosen as 100% of the signal intensity of sternocleidomastoid muscle on I2 (proton density-weighted reference image) which maximized the sum of sensitivity and specificity. (b) Applying this threshold identified IPH (red contour) on a SNAP image. (c) The matched histological section also showed IPH (blue contour) with a similar morphology.

Accuracy of IPH detection

With this optimal signal intensity threshold, the sensitivity and specificity for IPH detection at the slice level were 61% and 88% based on the histological assessment. IPH volume measured using the optimal threshold on SNAP and the IPH volume measured on histology were strongly correlated (Spearman’s rho = 0.76, p = 0.002).

Inter-reader agreement for manual detection of IPH on SNAP among the three readers was high (kappa: 0.92; 95% CI: 0.87, 0.96). Manual review detected IPH in 34 MR slices from 7 plaques. The corresponding estimates of slice-level sensitivity and specificity relative to histology were 63% and 96%, respectively. Both semi-automatic (kappa: 0.55; 95% CI: 0.34, 0.68) and manual (kappa: 0.62; 95% CI: 0.39, 0.77) review had moderate overall agreement with histology, whereas agreement between semi-automatic and manual review was high (kappa: 0.92; 95% CI: 0.83, 1.00).

Scan-Rescan Reproducibility of IPH Signals on SNAP

Maximum normalized intensity in SNAP showed high scan-rescan reproducibility with an ICC of 0.88 (95% CI: 0.74, 0.92) (Table 2). Using the histology-determined signal intensity threshold, 17 (26%) arteries were categorized as IPH-positive in both scans, 7 (11%) were categorized as IPH-positive in only one scan, and 42 (64%) were categorized as IPH-positive in neither scan, yielding a kappa of 0.75 (95% CI: 0.57, 0.91). The arteries showing IPH in only one scan were associated with a much smaller IPH volume compared to those showing IPH in both scans (3.0 mm³ ± 5.0 versus 34.9 mm³ ± 38.9; p<0.001).

Table 2.

Scan-Rescan Reproducibility of Quantitative Measures from Volumetric Image Processing.

	All Arteries (N = 66)			IPH-positive in both scans (N = 17)				IPH-positive in at least one scan (N = 24)
	ICC (95% CI)	CV	Within-subject SD	ICC (95% CI)	CV	Within-subject SD	p*	ICC (95% CI)	CV	Within-subject SD	p*
Maximum normalized intensity	0.88 (0.77, 0.93)	15.6%	0.19	0.86 (0.57,0.96)	10.6%	0.21	0.492	0.86 (0.70, 0.94)	12.8%	0.22	0.102
IPH volume (mm3)	-	-	-	0.99 (0.93, 1.00)	11.8%	4.13	0.062	0.99 (0.94, 1.00)	16.2%	4.06	0.022
Mean normalized intensity of IPH	-	-	-	0.77 (0.32,0.94)	5.2%	0.07	0.827	-	-	-	-

^*Paired t-test between scan 1 and scan 2.

- Measurements not obtainable or comparable in subjects without IPH.

CI = confidence interval, CV = coefficient of variation, ICC = intraclass coefficient variation, SD = standard deviation.

Of the 17 arteries categorized as IPH-positive in both scans, quantitative measures of IPH showed high reproducibility and small measurement variability (Table 2; Figure 3–4). The ICC statistic favored IPH volume while the CV statistic favored mean normalized intensity of IPH. Bland-Altman plots indicated no apparent relationship between variance and mean for both measurements (Figure 4).

Figure 3.

Automatic registration and segmentation of two SNAP scans of a 75-year old male. (a1, b1) Co-registered SNAP images are displayed using minimum intensity projection (top row), maximum intensity projection (middle row) and a combined view (bottom row). (a2, b2) Axial slices at four levels of IPH (yellow dashed lines in a1 and b1) compare segmented IPH regions (red contours) between the two scans.

Figure 4.

Scan-rescan reproducibility of quantitative measure of IPH signals. Scatter plots (left) and Bland-Altman plots (right) show the reproducibility of IPH volume (a) and mean normalized signal intensity (b) in the 17 arteries categorized as IPH-positive at both scans. Dashed lines in scatter plots are y = x lines.

Discussion

We describe a volumetric image processing method based on histology for quantitative characterization of carotid IPH using SNAP [29,30]. We derived a signal intensity threshold for IPH detection using IPH on histology as the basis for classifying a pixel as IPH on SNAP. Using this histology validated threshold, semi-automated detection of IPH showed good agreement with IPH detection by expert manual reviewers. Moreover IPH volume measured automatically using the threshold showed good accuracy compared to IPH volumes measured using histology. Furthermore, scan-rescan reproducibility of IPH detection and measurement was high using the semi-automated method.

Most previous studies have focused on the presence or absence of high T1 signals when studying high-risk plaques [8,9,16]. However, recent studies suggested that quantitative measures of T1 signals may serve as novel biomarkers for understanding the substantial heterogeneity among IPH plaques and monitoring IPH progression [10,13,31]. In a study of 31 patients with recent cerebrovascular ischemic events and bilateral carotid IPH, Wang et al. [13] found that IPH plaques on the symptomatic side had stronger and more extensively distributed T1 signals than those on the asymptomatic side. Both IPH progression and regression have been observed by measuring the extent or strength of T1 signals in 1–2 years [10–12]. Whether these changes reflect underlying tissue repair or repeated IPH remain unclear. In a serial study of coronary plaques, Noguchi et al [31] found that 12-month pitavastatin treatment resulted in a significant decrease in normalized T1 signal intensity compared to a significant increase in the control group. The quantitative, reproducible approach to IPH measurement as developed and validated in this study may provide additional information in such studies.

Few studies have reported reproducibility of IPH signal measurements on MRI. Using multicontrast MRI, Touze et al [32] obtained ICCs of 0.70 (95% CI: 0.52, 0.85) and 0.60 (95% CI: 0.37, 0.81) for intraobserver and interobserver IPH area measurement reproducibility, respectively. Inversion-recovery prepared gradient echo affords higher IPH-to-wall contrast [24], which may improve reproducibility of IPH signal measurements. By using phase-sensitive acquisition to increase the dynamic range of image contrast, SNAP further increases IPH-to-wall contrast and also reduces flow artifacts, which may further improve reproducibility of IPH measurement.

The SNAP images that were evaluated for IPH are inherently corrected for coil sensitivity. Since the SNAP image is calculated as a weighted ratio from acquired two images (I1 and I2) both of which share the same coil sensitivity as part of the phase sensitive reconstruction process, image intensity on the SNAP image does not suffer from coil sensitivity bias that occur in other sequences such as MPRAGE.

One previous study evaluated scan-rescan reproducibility of IPH signals on MRI [33]. A Pearson’s correlation coefficient of 0.97 was reported for IPH volume in 12 carotid arteries scanned twice within two weeks using a semi-automated segmentation algorithm. Compared to the previous study, the method described here does not require manual seed points of IPH for initialization. It works on 3D data directly and can thus handle tortuous arteries. Importantly, our method was developed and validated against histology rather than only manual review.

We found that both manual and automated IPH detection showed only a moderate agreement with histology, although high agreement between manual and automated IPH detection was achieved. These findings highlighted some intrinsic limitations of MRI in IPH measurement, including limited spatial resolution and possible co-localization of IPH with calcification, as previously seen for other sequences [24,28]. To ensure blinded analyses, matching between histological sections and reformatted SNAP images was performed indirectly by using 2D MR images as a bridge, which could also lower the agreement. Nonetheless, we expect that any mismatching should be random and not affect the signal intensity threshold.

Our study has the following limitations. 1) Compared to previous reproducibility studies [15–17], we adopted imaging inclusion criteria to enrich our study sample with carotid plaques. However, this study remains limited by the small number of IPH plaques. 2) A related limitation is that validation was done in a subset of patients. 3) The acquired resolution of SNAP scans (isotropic 0.8mm) may have reduced the agreement with histology especially in cases of small IPH. Improving the spatial resolution may improve IPH detection for small IPH.

In conclusion, a semi-automatic method based on volumetric processing of SNAP MRI for quantitative carotid IPH measurement was developed. The method was found to provide accurate IPH detection and reproducible measures of IPH volume, which may facilitate serial studies of carotid IPH.

Abbreviations

AUC

area under the curve

CEA

carotid endarterectomy

CI

confidence interval

CV

coefficient of variation

ICC

intra-class correlation coefficient

MRI

magnetic resonance imaging

ROI

region-of-interest

SNAP

simultaneous noncontrast angiography and intraplaque hemorrhage MRI

SCM

sternocleidomastoid muscle