Improved Left Ventricular Mass Quantification with Partial Voxel Interpolation – In-Vivo and Necropsy Validation of a Novel Cardiac MRI Segmentation Algorithm

Abstract

Background

CMR typically quantifies LV mass (LVM) via manual planimetry (MP), but this approach is time consuming and does not account for partial voxel components - myocardium admixed with blood in a single voxel. Automated segmentation (AS) can account for partial voxels, but this has not been used for LVM quantification. This study used automated CMR segmentation to test the influence of partial voxels on quantification of LVM.

Methods and Results

LVM was quantified by AS and MP in 126 consecutive patients and 10 laboratory animals undergoing CMR. AS yielded both partial voxel (AS_PV) and full voxel (AS_FV) measurements. Methods were independently compared to LVM quantified on echocardiography (echo) and an ex-vivo standard of LVM at necropsy. AS quantified LVM in all patients, yielding a 12-fold decrease in processing time vs. MP (0:21±0:04 vs. 4:18±1:02 min; p<0.001). AS_FV mass (136±35gm) was slightly lower than MP (139±35; Δ=3±9gm, p<0.001). Both methods yielded similar proportions of patients with LV remodeling (p=0.73) and hypertrophy (p=1.00). Regarding partial voxel segmentation, AS_PV yielded higher LVM (159±38gm) than MP (Δ=20±10gm) and AS_FV (Δ=23±6gm, both p<0.001), corresponding to relative increases of 14% and 17%. In multivariable analysis, magnitude of difference between AS_PV and AS_FV correlated with larger voxel size (partial r=0.37, p<0.001) even after controlling for LV chamber volume (r=.28, p=0.002) and total LVM (r=0.19, p=0.03). Among patients, AS_PV yielded better agreement with echo (Δ=20±25gm) than did AS_FV (Δ=43±24gm) or MP (Δ=40±22gm, both p<0.001). Among laboratory animals, AS_PV and ex-vivo results were similar (Δ=1±3gm, p=0.3), whereas AS_FV (6±3gm, p<0.001) and MP (4±5gm, p=0.02) yielded small but significant differences with LVM at necropsy.

Conclusions

Automated segmentation of myocardial partial voxels yields a 14-17% increase in LVM vs. full voxel segmentation, with increased differences correlated with lower spatial resolution. Partial voxel segmentation yields improved CMR agreement with echo and necropsy-verified LVM.

Introduction

Left ventricular mass (LVM) is widely used to guide clinical decision-making and prognostic assessment.^1,2 Cardiac magnetic resonance (CMR) is well suited to assess LVM, as it provides high-resolution tomographic imaging that enables volumetric quantification without geometric assumptions. LVM quantification on CMR is typically done using manual planimetry (MP), whereby myocardial borders are traced by hand. While MP has the potential to be highly detailed, it can be time consuming, especially when LV chamber dilation is present.³ MP can be especially challenging with respect to LV trabeculations, which contribute to LVM but are irregular in shape and can be difficult to discern from LV blood pool. Indeed, while MP of trabeculae has been reported to decrease CMR reproducibility and prolong MP processing time,^4,5 failure to account for trabeculae yields increased discordance with necropsy-verified LVM and alters clinical classifications for patients with LV remodeling, in whom trabecular size and complexity are increased.^6-8

Automated segmentation (AS) can rapidly quantify highly detailed structures within the heart and elsewhere. Recent AS advances have enabled quantification of partial voxels - discrete structures admixed within a single imaging voxel. This approach fundamentally differs from conventional (full voxel) analysis, whereby regions are partitioned in a binary manner based on location in relation to manual or automated contours. While new to CMR, partial voxel segmentation has been successfully applied to other organs: Neurologic studies have shown this to be useful for segmentation of tissue borders and irregularly contoured structures.^9-12 By extension, partial voxel segmentation holds particular relevance for LVM as it can account for endocardial irregularities and LV trabeculae.

AS algorithms have recently been developed that can simultaneously perform full and partial voxel segmentation of routine cine-CMR (SSFP) images.^13,14 Instead of generating shape-based contours that mimic MP, the algorithms measure statistics regarding signal intensities of blood and myocardium, and from this, perform voxel-by-voxel segmentation whereby partial voxel content of each substance is quantified for every LV voxel. In an initial validation study, partial voxel segmentation closely agreed with ex-vivo phantom volumes and was applied in-vivo for LV chamber quantification.¹³ However, to date, partial voxel segmentation has not been used to measure LVM.

This study tested LVM segmentation among clinical patients and laboratory animals undergoing CMR. In patients, echocardiography (echo) was performed within 1 day of CMR and used as a clinical comparator for LVM. In laboratory animals, sacrifice was performed after CMR and segmentation results were compared to ex-vivo LV weight. The aim was to examine the impact of partial voxel segmentation on CMR quantification of LVM.

Methods

Clinical Protocol

The study sample comprised consecutive patients enrolled in a post-myocardial infarction (MI) registry who underwent CMR and echo within a one-day interval.¹⁵ Imaging was performed between 9/2006 and 4/2010 at Weill Cornell Medical College. The study was conducted in accordance with the Weill Cornell Institutional Review Board (IRB); written informed consent was obtained at study enrollment.

CMR Image Acquisition

CMR was performed at 1.5 Tesla (General Electric) using a standard 2-dimensional steady state free precession (SSFP) pulse sequence. Short-axis SSFP images were acquired from the mitral annulus through the LV apex. Typical SSFP parameters were - repetition time 3.5 msec, echo time 1.6 msec, flip angle 60°, temporal resolution 30-50 msec, in-plane spatial resolution 1.6 mm × 1.3 mm. Images were acquired with a slice thickness of 6.0 mm and inter-slice gap of 4.0 mm. CMR imaging was performed without adjunctive contrast (gadolinium) administration.

LV Mass Quantification

LVM was quantified on CMR using MP and AS. AS segmentation was performed with simultaneous calculation of LVM based on full (AS_FV) and partial (AS_PV) voxel calculated myocardial content.

AS and MP were performed independently, blinded to results of the other method. For each, contiguous end-diastolic short axis images were segmented from LV base through apex. Basal and apical images of all patient exams were defined in accordance with established clinical criteria,^4,8 with the basal LV defined by the basal-most image encompassing at least 50% of circumferential myocardium. To directly compare AS to MP while minimizing potential confounding by slice and phase variance, basal-apical slice position and end-diastolic phase were held constant between methods. LVM was calculated as the product of myocardial specific gravity (1.05) and volume by each segmentation method.

Manual Planimetry

MP was performed in accordance with standard clinical practice, with LVM quantified by planimetry of end-diastolic endocardial and epicardial borders. Papillary muscles and trabeculae were included within myocardial contours; trabeculae were defined as myocardium protruding from the circumferential contour of the LV endocardium with similar signal intensity to the adjacent LV wall.^4,8 Planimetry was performed using commercial software (ReportCARD 4.0, General Electric). MP was also used to quantify LV dimensions, volumes, and total endocardial surface area – defined as the area of planimetered endocardial surface for each slice, summed across all 2D slices. Segmentation was performed by physicians (JWW, MDC) with ACC/AHA level III training in CMR.

Automated Segmentation

Automated LVM quantification was performed using integrated endocardial and epicardial segmentation based on previously established algorithms.^13,14 User input included identification of slices to be segmented and definition of the mitral and aortic valve annuli.

For endocardium, segmentation was performed using a geometry-independent algorithm that quantifies the mixture of blood and myocardium in each LV pixel.^13,14 Segmentation is accomplished by automatically computing blood and myocardial signal intensity distributions for each image individually (Figure 1A) and subsequently using that information to determine partial voxel content - defined as per voxel myocardial (or blood) content, for every voxel comprising the LV. The algorithm outputs two endocardial measurements: (1) “partial voxel” analysis consisting of the sum of partial blood voxel contents of every voxel (ENDO_PV), (2) “full voxel” analysis consisting of the sum of all voxels with any fractional blood content (ENDO_FV). Full voxel analysis closely mimics MP, which partitions myocardium and cavity in a binary manner, with trabeculae included in LVM when contiguous in shape or of similar signal intensity to surrounding LV myocardium.^4,7,8,16 Figures 1B and 1C provide an illustration of partial voxel content as quantified by the segmentation algorithm.

Figure 1. Automated Segmentation Algorithm.

A. Representative histogram generated by AS. Solid blue line = myocardial mean signal intensity. Dotted blue line = 2 standard deviations above myocardial mean (all intensities below this threshold considered full-myocardial voxels by AS). Solid green line: full-blood mean signal intensity. Dotted green line: 2 standard deviations below blood mean (all intensities above this threshold considered full-blood voxels). Blood and myocardial partial voxel content is linearly interpolated between dotted blue and dotted green lines. B. Endocardial voxel content and epicardial contour as calculated by AS. Endocardial voxel content is displayed on a voxel-by-voxel basis, with colors mapped from blue to green, which represents 100% myocardium (0% blood) to 100% blood (0% myocardium), respectively. Epicardial contour is displayed as a blue outline. C. Identical SSFP image with all endocardial voxels containing 100% blood removed from the display - only voxels containing less than 100% blood remain in the endocardial segmentation, and the epicardial contour is maintained as the outer outline.

For epicardium, segmentation was performed using an active contour model that uses location and signal intensity information resulting from the endocardial segmentation, in addition to signal intensity and edges at the epicardial-pericardial interface.¹⁴ LVM was calculated based on epicardial volume subtracted by either ENDO_PV or ENDO_FV respectively.

Intra- and inter-observer reproducibility of automated and manual CMR methods were tested among a random sample of 20 patients.

Validation Protocol

Each method – MP, as well as AS_FV and AS_PV - was compared to 2 standards for LVM:

Clinical Validation – Echocardiography

Echocardiography (echo) was performed within 1 day of CMR in all patients. LVM was quantified in accordance with established consensus guidelines:¹⁷ Linear measurements on 2D and M-mode echo were used to calculate LVM using a standard formula (0.8*{1.04[(LVIDd+PWTd+SWTd)³ – (LVIDd)³]}+0.6g) developed and validated based on necropsy-verified LV weight.^18-21 Echo analysis was performed blinded to CMR data.

2D and M-mode linear measurements by the designated, ACC/AHA level III certified, echo reader for the current study (RBD) have been previously shown to be closely correlated to each other in a series of 196 adults (r = 0.967, p< 0.001, mean Δ = 0.4 gm, SD = 10.2 gm; p = NS).²² Among 22 participants in another study,²³ LVM calculated from 2D linear measurements by the same reader (RBD) yielded values close to those obtained with use of M-mode recordings by a second reader (r = 0.94, P <.001; mean Δ = 0.9gm, SD = 9.5 gm, p = NS). Excellent reproducibility between measurements by a single experienced reader from separate echocardiograms has previously been shown in a series of 183 hypertensive adults for LVM (intra-class correlation coefficient [rho] = 0.93, P <.001, mean Δ = 1.7 g, SD = 18gm, p = NS) as well as LV chamber and wall dimensions (rho = 0.83 to 0.87) using echo methods applied in the current study.²⁴

Ex-Vivo Validation - Necropsy

Necropsy validation of LVM was obtained in a pre-existing cohort of animals that underwent CMR immediately prior to sacrifice, with confirmation of LVM based on ex-vivo weight.^6,25 Necropsy specimens were weighed within 30 minutes of animal sacrifice. For the current study, CMR images were retrieved from image archives, analyzed by MP and AS, and compared to necropsy-verified LV weight. To directly compare CMR methods to total LV weight at necropsy, segmentation of animal exams was performed with inclusion of all slices containing LV myocardium. Basal - apical slice position and segmentation phases were matched between CMR methods.

Statistical Methods

Continuous variables (expressed as mean±standard deviation) were compared using paired Student’s t-test for two-group comparisons. Categorical variables were compared using McNemar’s test for paired proportions. All continuous variables had qq-plots and histograms suggesting normality. Processing time and LV volume were compared using bivariate correlation coefficients and linear mixed-effects models. Multivariable linear regression analyses were used to evaluate associations between continuous variables.

Comparison of mean LVM by each quantification method was assessed using a linear mixed-effects model (taking into account the fact that echo, MP, AS_FV, AS_PV measurements are correlated within patients/animals) with an unconstrained covariance matrix. Multiple comparisons procedures were used to control for family-wise error rate: For patient data, adjustment was done using the Tukey-Kramer multiple comparison procedure, which enabled primary comparison to the reference of echo as well as relative differences between CMR methods. For animal data, CMR methods were compared to necropsy reference with adjustment using the Dunett-Hsu multiple comparison procedure. Two-sided p < 0.05 indicated statistical significance. Calculations were performed using SPSS 12.0 (SPSS Inc, Chicago, IL) and SAS 9.2 (SAS Inc, Cary, NC).

Results

Patient Sample

LVM segmentation was tested in 126 consecutive patients undergoing CMR as part of an ongoing study examining post-myocardial infarction LV remodeling.¹⁵ No patients were excluded based on clinical characteristics or image processing results. Table 1 details patient characteristics.

Table 1.

Patient Characteristics

Age (year)	57 ± 13
Male gender	81% (102)
Atherosclerosis Risk Factors
Hypertension	44% (55)
Hyperlipidemia	48% (61)
Diabetes Mellitus	20% (25)
Tobacco Use	34% (43)
Family History	27% (34)
Coronary Artery Disease History
Prior Myocardial Infarction	6% (7)
Prior Coronary Revascularization	10% (12)
Cardiovascular Medications
Beta-blocker	98% (124)
ACE Inhibitor/ARB	71% (89)
HMG-CoA Reductase Inhibitor	97% (122)
Aspirin	100% (126)
Thienopyridines	94% (119)
Myocardial Infarct Parameters
Infarct Related Artery
Left Anterior Descending	63% (80)
Right Coronary	29% (36)
Left Circumflex	8% (10)
Infarct Size (% myocardium)	17 ± 10
Post Myocardial Infarction Interval (days)	26 ± 8
LV Chamber Size and Function
Cardiovascular Magnetic Resonance
5/5/2013Ejection fraction (%)	51 ± 11
End-diastolic volume (ml)	154 ± 44
End-systolic volume (ml)	77 ± 37
Echocardiography
Ejection fraction (%)	48 ± 11
End-diastolic diameter (cm)	5.7 ± 0.5
End-systolic diameter (cm)	4.3 ± 0.6

Automated LV Mass Segmentation

AS successfully quantified LVM in all cases. Of the total 1127 images (126 exams, 8.9±0.9 images/exam), 51 (4.5%) required endocardial corrections, 74 (6.6%) epicardial corrections, and 107 (9.5%) delineation of the basal LV outflow tract. In aggregate, 60% of exams required no manual adjustment apart from identification of LV images and truncation of the LV outflow tract. For the remainder, epicardial or endocardial contours were manually adjusted based on visual inspection (1.0±1.6 adjustments per exam, 58% epicardial). Endocardial or epicardial corrections were deemed necessary when automated segmentation failed to truncate the myocardial border with either LV cavity or pericardium (i.e. due to temporal blurring and/or indistinct anatomic boundaries). Total processing time, including AS, visual inspection, and any manual adjustment, was under 1 minute in all cases.

Figure 2 provides a typical example of LVM segmentation by AS compared to MP.

Figure 2. Typical Example.

Typical short axis images demonstrating AS (2A) and MP (2B) segmentation, both of which account for myocardial trabeculae and papillary muscles. In this example, LV mass by AS_FV (75.0 g/m²) and MP (74.6 gm/m²) closely agreed, whereas AS_PV yielded higher mass (86.4 g/m²) respectively corresponding to relative differences of 15% and 16% with AS_FV and MP.

Full Voxel Segmentation

Table 2 reports LVM quantification by AS_FV and MP. LVM was slightly lower by AS_FV, with average differences between methods (3±9gm, 1.6±4.7gm/m², p<0.001) corresponding to 6±4% of LVM by MP. When established, CMR-based cutoffs were applied,^26,27 both methods yielded similar proportions of patients meeting criteria for LV hypertrophy (p=1.00) and chamber remodeling (p=0.73).

Table 2.

Conventional LV Mass Segmentation

LV Segmentation Results
	Manual Planimetry	Automated Segmentation (full voxel)	Δ	P
LV Mass (gm)	139.1 ± 35.1	135.9 ± 35.0	3.1 ± 9.3 gm	<0.001
LV Mass Index (gm/m2)	71.1 ± 15.1	69.5 ± 15.3	1.6 ± 4.7 gm/m2	<0.001
Diagnostic Classifications*
LV Chamber Remodeling	57% (72)	56% (70)	2% (2)	0.73
LV Hypertrophy	3% (4)	4% (5)	1% (1)	1.00

Processing Time
	Manual Planimetry	Automated Segmentation	P
Processing Time	4:18 ± 1:02	0:21 ± 0:04	<0.001

Bold face type indicates p value < 0.05 (data presented as mean±SD)

^*Based on previously established, population-based, CMR cutoffs^26,27

Also shown in Table 2, processing time for AS was over 12-fold lower vs. MP, corresponding to an average time savings of nearly 4 minutes (Δ=3:58±1:02, p<0.001). Processing time correlated with LV chamber size for MP (r=0.57, p<0.001) and AS (r=0.24, p=0.007). However, as shown in corresponding scatterplots for each method, regression slopes were over 30-fold higher for MP (0.89) compared to AS (.03), reflecting a markedly higher proportionate increase in processing time in relation to chamber size.

Partial Voxel Segmentation

AS was also used to calculate LVM with incorporation of myocardial partial voxels (AS_PV). Table 3 compares AS_PV to MP and AS_FV. As shown, LVM by AS_PV was higher compared to either MP (Δ= 20±10 gm) or AS_FV (Δ= 23±6 gm, p<0.001), corresponding to relative differences of 14% and 17% respectively. LVM by AS_PV yielded a similar proportion of patients (5%) meeting established criteria for LV hypertrophy²⁶ compared to either MP (3%; p=0.50) or AS_FV (4%; p=1.00). However, AS_PV yielded a markedly lower proportion of patients (26%) meeting established criteria for LV chamber remodeling²⁷ than did MP (57%; p<0.001) or AS_FV (56%; p<0.001).

Table 3.

Comparison of Full and Partial Voxel Adjusted LV Mass

	Partial Voxel
	ASPV	ASFV			MP
	Mean ±SD	Mean ±SD	Δ*	p*	Mean ±SD	Δ*	p*
LV Mass (gm)	158.7 ± 37.9	135.9 ± 34.9	22.8 ± 5.5	< 0.001	139.1 ± 35.1	19.7 ± 10.1	< 0.001
LV Mass Index (gm/m2)	81.2 ± 16.2	69.5 ± 15.3	11.7 ± 2.4	< 0.001	71.1 ± 15.1	10.1 ± 5.0	< 0.001

^*difference (mean±SD) vs. partial voxel adjusted LV mass (p values adjusted for multiple comparisons)

In multivariable analysis, magnitude of difference between AS_PV and AS_FV independently correlated with larger voxel size (partial r=0.37, p<0.001) even after controlling for LV chamber volume (r=0.28, p=0.002) and LVM (r=0.19, p=0.03) as quantified by the standard of full voxel segmentation (model r=0.64, p<0.001).

All segmentation methods demonstrated good intra- and inter-reader reproducibility, although limits of agreement were smaller for both AS_FV and AS_PV by AS compared to MP (Figure 3).

Figure 3. Reproducibility.

Bland-Altman plots demonstrating reproducibility data for each segmentation method (data shown as mean±2 standard deviations). 3A provides intra-observer data, demonstrating that mean differences for all methods were small (MP: 1.1gm, AS_FV: -0.3gm, AS_PV -0.6gm), although limits of agreement were narrower for AS_FV (-4.3gm to 3.8gm) and AS_PV (-5.6gm to 4.5gm) compared to MP (-12.1gm to 14.3gm). Inter-observer measurements (3B) demonstrated similar findings.

Validation

Each CMR segmentation method was independently compared to two standards; (1) a clinical standard of LVM measured on echo, and (2) an ex-vivo standard of LVM as weighed at time of necropsy.

LV Mass by Echocardiography

Echo was performed within 1 day of CMR in all patients (97% same day); 96% of echoes (n=121) were technically sufficient to quantify LVM. The most common reasons for technically insufficient echoes (4%) were poor endocardial definition or off-axis imaging.

Figure 4A shows LVM results by each method, demonstrating that echo yielded higher LVM than did all CMR segmentation methods (p<0.001). However, as shown in 4B, mean differences between CMR and echo were smallest with LVM_PV (Δ= 20±25gm, 11±13gm/m²) compared to LVM_FV by either AS (Δ= 43±24gm, 22±12gm/m²) or MP (Δ= 40±22gm, 20±12gm/m²) (both p<0.001).

Figure 4. Comparison to Echocardiography.

A. Mean LVM by each CMR segmentation method (gray bars) compared to echo (black bar) among patient cohort (n=121).

B. Mean LVM difference between each CMR method and echo, demonstrating smaller differences with AS_PV as compared to AS_FV or MP (p<0.001; all p values adjusted for multiple comparisons)

LV Weight at Necropsy

LVM segmentation methods were also tested in a pre-existing cohort of 10 animals (8 dogs, 2 pigs) that underwent CMR prior to sacrifice. Figure 5 shows results of each CMR method compared to the reference standard of ex-vivo LV weight: LVM averaged 70±13gm at necropsy and 69±14gm by LVM_PV, reflecting non-significant absolute and relative differences of 1±3gm and 4±3% (p=0.54). Automated LVM_FV was lower (63±14gm) and yielded significant differences (7±3gm, 10±6%) with ex-vivo LV weight (p<0.001). MP paralleled automated full voxel data, as reflected by lower LVM (65±16gm) that differed significantly (4±5gm, 8±7%; p=0.049) from ex-vivo results.

Figure 5. Comparison to Necropsy.

LVM by each CMR segmentation method (gray bars) compared to necropsy derived LV weight (black bar) among animal cohort (n=10). Only AS_PV yielded non-significant differences with LV weight at necropsy (all p values adjusted for multiple comparisons).

Discussion

This study is the first to examine partial voxel segmentation for automated LVM quantification. There are several key findings: First, among the consecutive series of post-MI patients studied, AS using full voxel analysis (AS_FV) yielded similar results to manual planimetry (MP), with small, albeit statistically significant, absolute differences (3±9gm, p<0.001). Both methods yielded similar results concerning classification of patients with LV remodeling (p=0.73) or hypertrophy (p=1.00). Second, AS using partial voxel analysis (AS_PV) yielded larger LVM than AS using full voxel analysis (Δ=23±6gm) or MP (Δ=20±10gm; both p<0.001). Magnitude of difference between AS_PV and AS_FV independently correlated with larger voxel size (partial r =0.37, p<0.001) even after controlling for LV chamber volume (r=0.28, p=0.002) and LVM (r=0.19, p=0.03) (model r=0.64, p<0.001). Third, AS_PV yielded better agreement with the clinical standard of LVM by echo and smaller differences with LV weight at necropsy.

Regarding necropsy data, automated segmentation was applied to a pre-existing CMR dataset in which actual LV weight was verified ex-vivo. Results demonstrated that partial voxel segmentation yielded non-significant differences with necropsy-evidenced LVM (1±3gm, p=0.3) whereas full voxel yielded small but significant differences when either AS (6±3gm, p<0.001) or MP (4±5gm, p=0.02) were used. This finding is consistent with prior cine-CMR (SSFP) studies that have reported small mean differences between CMR and necropsy LVM, but have noted variability ranging from -0.8±2.6 to 0.2±8.4gm with MP, and -10.6±7.1 to 4.2±7.1gm with AS.^6,28,29 While the reasons for variable differences in CMR and necropsy results are not certain, this may relate to animal and study specific differences in spatial resolution, image quality, or interval between imaging and necropsy with resultant post-mortem changes in LVM.

To further test CMR segmentation vs. an independent clinical reference, patient results were compared to LVM as quantified by echo. Findings demonstrated that whereas all CMR methods yielded lower LVM than echo, partial voxel analysis yielded smaller mean differences (20±25gm, 11±13gm/m², p<0.001) than did AS full voxel (43±24gm, 22±12gm/m²) or MP (40±22gm, 20±12gm/m²). Prior comparative studies have used MP and also reported lower LVM by CMR,^30-32 although variance has been larger than in our study, as evidenced by mean differences of 37±39gm/m² in cohorts with valvular heart disease and 58±63gm (data reported un-indexed) in heart transplant patients.^31,33

Our data shed new light on prior comparative studies, which have commonly attributed LVM differences between modalities to echo-specific factors. Certainly, it is important to recognize that echo-based calculations employ geometric assumptions whereas CMR quantifies LVM based on actual planimetry of myocardial borders: Both off axis imaging and image quality can affect echo measurements, whereas CMR provides high resolution imaging with great precision for detecting small differences in LVM.³³ However, prior papers have demonstrated systematically lower LVM by CMR vs. echo,^30-32 suggesting intrinsic biases not fully explained by echo alone in context of several studies showing unbiased estimation of echo-derived LVM vs. necropsy-verified LV weight.^18,19,21

Our results suggest that previously reported differences between CMR and echo may be partially attributable to the approach used for CMR segmentation, with agreement improved through quantification of myocardial partial voxel content. While reasons for improved agreement between echo and partial voxel CMR are uncertain, we speculate that this may be due to the fact that echo-formulae use linear measurements to calculate LVM based on models derived from actual LV weight (i.e. trabeculae inclusive) at necropsy, ^18,19,21 an approach that can yield error on an individual patient basis but provide generally accurate LVM when measured for overall populations. Partial voxel CMR calculates LVM without geometric assumptions while accounting for detailed components of LV myocardium (i.e. trabeculae) that can be difficult to trace manually but contribute to overall LV weight, resulting in higher LVM values for individual patients and across populations. These issues may explain improved agreement between partial voxel CMR and echo, as well as residual differences between modalities. Consistent with this, our group’s prior research has demonstrated that failure to segment trabecular volume on CMR yields increased discrepancy with linear echo formulas for LVM.⁸ As multimodality imaging is increasingly being used to guide patient care, the ability of partial voxel CMR segmentation to yield improved agreement with echo-derived LVM is of substantial clinical importance.

Beyond partial voxel segmentation, a novel feature of the AS algorithm tested in this study is that no geometric assumptions regarding endocardial shape are employed.¹³ This differs from several prior AS algorithms, which have employed shape based constraints regarding LV border geometry.^34-38 In contradistinction, the algorithm tested in this study relies on only one fundamental assumption - LV blood is enclosed by LV myocardium. Based on this, AS is performed using an automated effusion-threshold based approach that relies on intrinsic differences in signal intensity between blood and myocardium rather than shape-based algorithmic constraints. This feature is complementary to partial voxel segmentation, in that it enables LV segmentation independent of remodeling-associated changes in LV contours. Moreover, the current algorithm segmented all cases in less than 1 minute, a considerable time saving compared to MP. Whereas few prior studies have reported actual processing times, we note that mean processing time for AS in the current study (0:21±0:04) was far lower than that reported for prior shape-based AS algorithms (5:00±0:18 minutes).²⁹

Concerning clinical performance, study results demonstrate the utility of geometry independent effusion-threshold segmentation. Among the consecutive series of 126 patients tested, AS was successful in all cases and required minimal user corrections. Absolute differences between AS_FV and MP were significant but small (Δ=3±9gm, p<0.001), resulting in non-significant differences when established CMR criteria for LV hypertrophy and remodeling were applied. The relative agreement between MP and AS_FV can be explained at least in part by the similarities between the two techniques. When performing MP, an operator visually discerns LV myocardium from blood based on shape or signal intensity,^4,7,8,16 with regions of similar signal intensity partitioned together. With this approach, voxels that contain any amount of fractional blood content (and thus exhibit higher signal intensity than the adjacent myocardium) are included in blood volume, even though they may also include fractional myocardial content. Similarly, AS_FV employs an algorithm that simply counts the number of voxels with any fractional blood content - thus, it labels regions that are inherently brighter than LV myocardium (due to fractional blood content) as blood, even though they may also contain fractional myocardial content. In contrast, partial voxel segmentation allows fractional quantifications of both blood and myocardium within each voxel region. Thus, AS_PV would be expected to be of incremental utility when increased voxel size results in juxtaposition of myocardium and blood within a single voxel - a phenomenon that is not accounted for by MP or full voxel analysis. Consistent with this, our results demonstrate that differences between full and partial voxel segmentation correlate with larger voxel size (r=0.37, p<0.001) even after controlling for LV cavity size and total LVM.

Several limitations should be recognized. First, clinical performance of CMR segmentation was tested among patients in a post-MI registry, the majority (81%) of whom were male. While this enabled us to test segmentation among patients with MI-associated LV remodeling, further study is needed to evaluate partial voxel segmentation in broad population-based cohorts. Second, although results demonstrate that partial voxel segmentation yields better agreement with independent standards of echo and necropsy quantified LVM, clinical outcomes data were not obtained and the predictive value of partial vs. full voxel CMR segmentation results is not known. Finally, while geometry-independent partial voxel segmentation was shown to perform robustly, minimal user interaction is still required to identify the actual LV slices to be segmented.

In summary, this study demonstrates that partial voxel automated segmentation is a promising improvement for left ventricular mass quantification. Future research is necessary to test whether partial voxel adjusted LVM provides incremental utility vs. full voxel assessment for clinical prognostic assessment.

Clinical Summary.

Left ventricular mass (LVM) is widely used to guide clinical decision-making and prognostic assessment. Cardiac magnetic resonance (CMR) is well suited to assess LVM, as it provides high-resolution tomographic imaging that enables volumetric quantification without geometric assumptions. CMR typically quantifies LVM by manual planimetry (MP) of LV chamber contours: This approach is widely used in clinical practice but can be time consuming, challenging with respect to planimetry of irregularly contoured trabeculae, and limited in its ability to account for myocardial partial voxels – myocardium admixed with blood in a single voxel. Recent advances in automated segmentation (AS) have enabled quantification of partial voxel components. This study tested a novel AS algorithm that can quantify LVM while accounting for myocardial partial voxels. Among laboratory animals undergoing CMR prior to sacrifice, LVM quantified using AS based partial voxel segmentation (AS_PV) yielded non-significant differences with LV weight at necropsy, whereas both conventional AS and MP yielded small but significant underestimations. Among patients undergoing CMR within 1 day of echocardiography, AS_PV yielded significantly smaller differences with echocardiography-quantified LVM than did either conventional AS or MP. AS was successful in all patients, required minimal manual adjustments (1.0±1.6 per exam), and yielded a 12-fold reduction in processing time vs. MP (0:21±0:04 vs. 4:18±1:02 min; p<0.001). These results support use of automated partial voxel segmentation for quantification of LVM, demonstrating that this method markedly reduces CMR processing time among clinical patients, while yielding improved agreement with independent references of both echocardiography and necropsy-verified LVM.

Common Abbreviations

CMR

cardiac magnetic resonance

MP

manual planimetry

AS_FV

automated segmentation using full voxel analysis

AS_PV

automated segmentation using partial voxel analysis

LVM

left ventricular mass