Correcting posterior paraspinal muscle CT density for intravenous contrast material independent of sex and vascular phase

Abstract

Purpose.

Intravenous contrast poses challenges to CT muscle density analysis. We developed and tested corrections for contrast-enhanced CT muscle density to improve muscle analysis and utility of CT scans for the assessment of myosteatosis.

Materials and Methods.

Using retrospective images from 240 adults who received routine abdominal CT imaging from March to November 2020 with weight-based iodine contrast, we obtained paraspinal muscle density measurements from non-contrast, arterial, and venous phase images. We used a calibration sample to develop 9 different mean and regression-based corrections for the effect of contrast. We applied the corrections in a validation sample and conducted equivalence testing.

Results.

We evaluated 140 patients (mean age 52.0 years [SD:18.3]; 60% female) in the calibration sample and 100 patients (mean age 54.8 years [SD:18.9]; 60% female) in the validation sample. Contrast-enhanced muscle density was higher than non-contrast by 8.6 HU (SD:6.2) for the arterial phase (female, 10.4 HU [SD:5.7]; male, 6.0 HU [SD:6.0]) and by 6.4 HU [SD:8.1] for the venous phase (female, 8.0 HU [SD:8.6]; male, 4.0 HU [SD:6.6]). Corrected contrast-enhanced and non-contrast muscle density was equivalent within 3 HU for all corrections. The −7.5 HU correction, independent of sex and phase, performed well for arterial (95% CI: −0.18, 1.80 HU) and venous phase data (95% CI: −0.88, 1.41 HU).

Conclusions.

Our validated correction factor of −7.5 HU renders contrast-enhanced muscle density statistically similar to non-contrast density and is a feasible rule-of-thumb for clinicians to implement.

Introduction

The high resolution and reproducibility of CT images can be used for the analysis of muscle density, which has been proposed as a potential diagnostic tool for muscle wasting disorders such as sarcopenia, cachexia, and frailty.¹ In fact, low muscle density is predictive of mortality in numerous populations.^2–6 Over 90 million CT scans are ordered in North America alone⁷ which contain extensive health data beyond their clinical indications, however, muscle density analysis of CT exams is not part of standard medical care.

One underappreciated barrier to utilizing CT scans for muscle analysis is the effect of intravenous (IV) contrast material on muscle density, measured in Hounsfield Units (HU).¹ Contrast material perfuses into muscle and increases x-ray attenuation, which artificially increases muscle density^8–10 yet a recent review found that 94% of studies analyzing CT scans for muscle density did not report presence or absence of contrast in their scanning protocol.¹ It is also known from prior studies that contrast can artificially inflate muscle density¹¹ and incorrectly diagnose a patient as healthy although they have features of muscle wasting. Un-corrected factors that influence CT numbers restricts the utility of opportunistic CT scan analysis and impedes the development of standardized cut-points for healthy muscle density.^12,13

While contrast generally increases muscle density, the magnitude of contrast enhancement varies between studies.^11,14 This discrepancy may be partially explained by the variation in timing between contrast injection and CT image acquisition,^15,16 since the degree of enhancement is dependent on perfusion time.¹⁷ These discrepancies in CT contrast phases and protocols further complicate our ability to standardize muscle density measurements.

Therefore, the purpose of our study was to characterize the effects of contrast enhancement at different phases and to determine if a clinically feasible correction factor can adjust contrast-enhanced muscle density to non-enhanced muscle density. Determining how to harmonize contrast and non-contrast scans will be crucial to effectively utilizing opportunistic data in CT exams, including developing density-based cut-points of healthy muscle and applying those cut-points to opportunistic CT scans for diagnostic purposes. This research will improve the clinical and research utility of CT data irrespective of variations in CT protocol, including prediction of patient survival and assessment of myosteatosis and other muscle wasting disorders.

Materials and Methods

CT image identification

We obtained retrospective CT images from the Picture Archiving and Communication System (PACS) with a waiver of consent approved by the Institutional Review Board. Consecutive patients from a single location with a routine contrast abdominal CT exam and with a weight-based contrast administration protocol from March 2020 to November 2020 were included (n=240, Figure 1). Exam protocol included bolus-tracking, allowing for acquisition of baseline non-contrast images (NC), late arterial phase contrast images (AC), and venous phase contrast images (VC).¹⁸ We selected the first 140 patients from our cohort for the calibration sample, from which we developed mean-derived and regression corrections. We tested the correction factors on the validation sample of the remaining 100 patients (Table 1).

Figure 1.

Study flow chart.

Table 1.

Muscle density comparisons between calibration sample and validation sample

	Calibration sample			Validation sample			Comparison between samples (p-values)
	Female	Male	All patients	Female	Male	All patients	F vs F	M vs M	All
N	84 (60%)	56 (40%)	140	60 (60%)	40 (40%)	100	–	–	–
Age (years)	52.1 (18.2)	51.8 (18.5)	52.0 (18.3)	54.0 (18.2)	55.8 (20.0)	54.8 (18.9)	0.535	0.319	0.259
BMI	26.5 (4.4)	26.9 (4.1)	26.6 (4.3)	27.1 (4.5)	27.4 (3.1)	27.2 (4.0)	0.433	0.543	0.321
Baseline density (HU)	48.0 (7.7)	50.9 (15.0)	49.1 (11.2)	46.1 (11.9)	48.8 (11.3)	47.2 (11.7)	0.305	0.437	0.203
Arterial density (HU)	58.3 (8.2)	56.9 (16.1)	57.8 (12.0)	55.6 (12.5)	55.5 (11.6)	55.5 (12.0)	0.138	0.614	0.159
Venous density (HU)	55.9 (9.3)	54.9 (17.6)	55.5 (13.2)	55.3 (13.6)	54.5 (10.6)	55.0 (12.5)	0.757	0.890	0.748
Average density (HU)	57.1 (8.0)	55.9 (16.6)	56.6 (12.2)	55.4 (12.7)	55.0 (10.8)	55.3 (12.0)	0.365	0.745	0.382

All data displayed in means with standard deviations in parentheses except for N, which is displayed in number and percent in parentheses. P-values represent t-tests between female calibration vs. female validation sample (F vs. F), male calibration vs. male validation sample (M vs. M), and all patient calibration vs. all patient validation sample (All). HU, Hounsfield Unit.

Patients were scanned on a GE Discovery HD750 (GE Healthcare, Waukesha, WI). Technical factors for VC included: 120kV, helical pitch of 0.516:1, detector collimation of 64×0.625 mm, 3.75 mm slice thickness, noise index = 18.5, reconstruction kernel of “STANDARD”. NC and AC were acquired using a bolus tracking protocol at 120 kV and in 24 mAs in axial mode. Patients were given a weight-adjusted dose of Iohexol (350 mg/mL) at 3 mL/sec with a 50 mL saline flush. We selected medium sized patients that ranged from 58.97 kg to 113.40 kg (130 to 250 lbs) whose contrast dose ranged linearly from 80 mL to 150mL. ¹⁸ Reasons for routine abdominal CT exams were: 65% abdominal pain, 6% evaluation for diverticulitis, 5% hernia, 3% infection, and 21% other indications.

All images obtained were at the level of the liver, ranging from T9 to L3 (2% T9, 12% T10, 33% T11, 36% T12, 14% L1, 1% L2, 1% L3). One lower mA axial image was taken before contrast injection and used to obtain NC.¹⁸ Forty seconds after injection, a bolus-tracking protocol (SmartPrep©, GE) ran until a region of interest (ROI) in the liver reached 50 HU (Figure 2), initiating the venous phase acquisition.¹⁸ We used the first bolus-tracking image for the AC. Finally, we used the corresponding axial image from the full-abdominal venous phase CT for the VC (60 to 80 seconds after injection).

Figure 2.

Muscle density measurements and demonstration of density increase due to contrast. Red ovals represent region of interest measurements in the erector spinae muscles from non-contrast baseline (A), arterial phase contrast (B), and venous phase contrast (C) CT images. The average of left and right regions of interest muscle density measurements are show for each image. HU, Hounsfield Units.

CT skeletal muscle analysis

A single trained researcher (5 years training) conducted CT image analysis in Osirix MD (Pixmeo, Bernex, Switzerland). A 2 cm² ovular ROI was placed on both the left and right erector spinae muscles (if less than 2cm², then largest possible) and averaged to determine skeletal muscle density in HU (Figure 2). The researcher matched the size and location of ROIs at NC to the AC and VC (Table 2).

Table 2.

Calibration sample summary of muscle densities between phases.

	Baseline density	Arterial density	Venous density	Average arterial + venous density
All patients	49.1	57.8	55.5	56.6
Female	48.0	58.3	55.9	57.1
Male	50.9	56.9	54.9	55.9
F v M p-value	0.175	0.555	0.693	0.617

Differences between baseline and contrast phases were used to create mean corrections. Baseline, arterial, and venous, are significantly different (p<0.05) within all patients and both sexes. Female versus male densities were not different within non-contrast and phase densities. All muscle density values represented as Hounsfield Units.

Statistical Methods: Correction development and calibration sample

Statistics were performed in R Version 4.0 (R Foundation for Statistical Computing, Vienna, Austria). We collected data on 140 subjects to give a sufficiently narrow confidence interval around the estimated CT muscle HU correction. Based on previously collected pilot data, we assumed a SD of the population mean correction estimate to be 9 units. With an SD of 9 and N = 140, a 95% CI width of the correction estimate would be +/− 1.5 units. We assessed for demographic differences between calibration and validation samples with t-test or chi-square test for continuous or categorical variables, respectively with α=0.05.

We developed 9 corrections from the calibration sample of 140 subjects: 6 mean-derived, univariate corrections and 3 multivariate regression corrections. Three mean-derived, univariate corrections were developed by calculating the difference between NC muscle density and either the AC muscle density, VC density, or the average density (average of AC and VC density) for the calibration sample (Table 3). To develop sex-specific mean-derived, univariate corrections, we stratified the calibration sample by sex and calculated the same differences between NC muscle density and each phase, finding different mean correction values for males and females. When applying sex-specific correction factors to estimate the true HU, the correction factor specific to females and the correction factor specific to males are applied and then analysis was conducted on the full data set, not separately by sex. Finally, we developed 3 multivariate regression corrections using multiple linear regression to predict non-contrast muscle density with AC density, VC density, and the average density as independent variables with age, body mass index, and sex as covariates (Table 4). For both the mean and regression corrections, the AC corrections were applied to the AC validation data, the VC corrections were applied to the VC validation data, and the average corrections were applied to both AC and VC validation data (Table 5).

Table 3.

Mean corrections developed from calibration sample.

	Difference Baseline-Arterial	Difference Baseline-Venous	Difference Baseline- Average
All patients	−8.6	−6.4	−7.5
Female	−10.4	−8.0	−9.2
Male	−6.0	−4.0	−5.0

Sex-specific corrections were applied simultaneously in the analysis. All muscle density differences represented as Hounsfield Units.

Table 4.

Calibration sample regression models to correct for contrast.

CT feature or patient characteristic	Arterial phase density		Venous phase density		Average arterial + venous density
	Estimate	P-value	Estimate	P-value	Estimate	P-value
Intercept	2.73	0.538	8.68	0.092	1.02	0.822
Post HU	0.78	< 0.001	0.63	< 0.001	0.76	< 0.001
Age	−0.05	0.072	−0.11	0.001	−0.06	0.029
BMI	0.10	0.343	0.37	0.004	0.26	0.02
Male	3.98	< 0.001	3.41	0.002	3.75	< 0.001
R2	0.777		0.697		0.776

Regression models were developed to correct arterial phase muscle density, venous phase density, and an average of venous and arterial phase density to simulate disregarding contrast timing. HU, Hounsfield Unit.

Table 5.

Equivalence testing between corrected contrast density and non-contrast density in validation sample using 3 HU as margin of equivalence.

Correction Method	Difference between corrected contrast density and non-contrast density
	Correction applied to arterial data	Correction applied to venous data
Sex stratified mean correction
Arterial	−0.30 (−1.14, 1.47)	–
Venous	–	1.38 (0.29, 2.48)
Averaged phase correction	0.81 (−0.15, 1.77)	0.27 (−0.83, 1.36)
Non-sex stratified mean correction
Arterial	−0.30 (−1.29, 0.68)	–
Venous	–	1.38 (0.24, 2.53)
Average phase correction	0.81 (−0.18, 1.80)	0.27 (−0.88, 1.41)
Regression Correction
Arterial	0.10 (−0.83, 1.03)	–
Venous	–	1.47 (0.45, 2.49)
Average phase correction	1.04 (0.08, 2.00)	0.62 (−0.33, 1.57)

Each correction was either applied to arterial data (middle column), venous data, (right column), or both. Differences between corrected contrast scan densities and non-contrast scan densities are shown in HU with 95% confidence intervals in parentheses. Confidence intervals within the 3 HU threshold represent statistical equivalence between corrected contrast and non-contrast densities. Boxes marked with a dash did not have correction factors applied to the data in that column. HU, Hounsfield Unit.

Statistical Methods: Equivalence testing on validation sample

Prior to the study, we assumed an equivalence margin of 3 HU for the difference in the error of the corrected values’ estimation of the true HU. We tested all corrections with a validation sample of 100 patients. Based on a paired equivalence test with α=0.05, a 3-unit margin of equivalence, and assuming the true mean error is 0, 100 subjects gave us 90% power with a variability estimate of the error of SD=9 units for a two one-sided tests (TOST) of equivalence. We applied the 3 mean-derived non-sex specific corrections, 3 mean-derived sex-specific corrections, and 3 regression corrections to the contrast-enhanced muscle density values for the corresponding AC, VC, or average of the phases.

To compare each participant’s corrected contrast-enhanced muscle density value and NC value, we used TOST equivalence testing with a threshold of +/−3 HU indicating equivalence with α=0.05. We chose the +/−3 HU threshold to match a reasonable degree of intra-individual CT scan variation.

Statistical Methods: Difference between correction errors in the validation data

To determine if the correction methods produced statistically different correction errors from one another, differences between correction errors from mean and regression corrections were compared with paired t-tests.

Results

Correction development and calibration sample

The mean NC muscle density for the calibration sample was 49.1 (±11.2) HU, AC density was 57.8 (±12.0) HU, VC density was 55.5 (±13.2) HU, and averaged phase density was 56.6 (±12.2) HU (Table 1). We found no significant differences in sex, age, BMI, or muscle density at any phase between the calibration and validation samples. In our calibration sample, the mean muscle density for the NC, AC, and VC groups were all statistically different from each other (Table 2). Hence, the difference between contrast and NC muscle density, i.e. correction value, was also significantly different at each phase. There was no significant difference in male or female muscle density values at any phase. However, females had greater nominal differences between contrast-enhanced and NC muscle density, i.e. larger correction values, than males at each phase (p<0.01). Our sample contained both oral+IV contrast patients and IV contrast-only patients, however, we found no significant difference in muscle HU between these groups, so they were grouped together (data not shown). Thirty-eight patients were excluded for fixed-contrast dosing, and 5 were excluded for missing images or unreadable CT images. The selection of participants is shown in Figure 1.

Equivalence testing on validation sample

Equivalence testing revealed that all mean and regression corrections corrected contrast-enhanced values to be not significantly different from non-contrast values with a threshold of 3 HU (p<0.01).

AC muscle density in the validation sample was higher than NC muscle density by 9.5 ± 5.1 HU, 6.7 ± 4.3 HU, and 8.3 ± 5.0 HU for female, male, and total subjects, respectively (p<0.001; Table 1). Applying the mean-derived arterial corrections from the calibration sample (mean correction: −8.7 HU; sex-specific mean corrections: female −10.3 HU, male −6.0 HU; Table 3) to arterial phase data in the validation sample yielded an average muscle density difference of −0.30 HU (Table 5). Equivalence testing revealed that mean corrected AC values and NC values were not significantly different with a threshold of 3 HU (p<0.001).

VC muscle density in the validation sample was higher than NC muscle density by 9.2 ± 6.1 HU, 5.7 ± 4.5 HU, and 7.8 ± 5.8 HU for female, male, and total subjects, respectively (p<0.001; Table 1). Applying the mean-derived venous phase corrections from the calibration sample (mean correction: −6.4 HU; sex-specific corrections: female −7.9 HU, male −4.0 HU; Table 3) on venous phase data from the validation sample yielded an average 1.38 HU difference. Equivalence testing revealed that mean corrected VC values and NC values were not significantly different with a threshold of 3 HU (p<0.01).

When taking the average of AC and VC muscle density in the validation sample (to reduce the influence of contrast timing), contrast muscle density was higher than the NC density by 9.3 ± 4.9 HU, 6.2 ± 3.7 HU, and 8.0 ± 4.7 HU for female, male, and total subjects, respectively (p<0.001; Table 1). Applying the average-phase, mean-derived arterial and venous correction from the calibration sample (mean correction: −7.5 HU; sex-specific corrections: female −9.1 HU, male −5.0 HU; Table 3) to arterial phase data from the validation sample yielded an average 0.81 HU difference in muscle density. Applying this average correction to venous phase data in the validation sample yielded an average −0.27 HU difference in muscle density between NC and corrected VC density when using sex-specific corrections and when ignoring sex (Table 5). Equivalence testing revealed that mean corrected arterial and venous phase values and NC values were not significantly different with a threshold of 3 HU (p<0.001). The relationship of non-contrast data to contrast and corrected contrast data can be seen in Figure 3.

Figure 3.

Scatterplot of non-contrast vs. contrast erector spinae muscle density and corrected contrast muscle density in HU. Arterial (A) and venous (B) data show linear relationships (R²=0.74 and 0.63, respectively). Diagonal line represents perfect agreement between noncontrast and contrast data.

When using an arterial-phase regression correction developed from the calibration sample on the validation sample (Table 4), the AC density was corrected to a 0.10 HU difference as compared to the NC density (Table 5). Similarly, when using a venous-phase regression correction, the VC density was corrected to an average 1.47 HU difference to the NC density. When using a regression correction that averaged the arterial and venous calibration data, the AC density of the validation sample was corrected to a 1.04 HU difference to the NC density and the VC density was corrected to a 0.62 HU difference to the NC density. All regression corrections were less than the equivalence test threshold of 3 HU, meaning the corrected contrast density was not statistically different on average from the true non-contrast density (p<0.01).

Difference between correction errors in the validation data

There was no significant difference between the correction error in the validation data when arterial mean and arterial regression corrections (−0.30 vs 0.10; p=0.197) and venous mean and venous regression corrections (1.38 vs 1.47; p=0.864) were applied to the validation sample. Similarly, no differences were seen when average mean and regression corrections were applied to either arterial (0.81 vs 1.04; p=0.504) or venous phase (0.27 vs 0.62; p=0.336) data.

Discussion

We aimed to develop a correction method to adjust contrast-enhanced muscle density to density values in non-contrast images. The purpose of this correction is to improve the utility of CT exams for opportunistic muscle analysis when there is heterogeneity in CT scan protocols that can modulate muscle density measurements. We found that all 9 of our corrections were able to adjust contrast-enhanced muscle density to be within 3 HU of non-contrast muscle density. The most clinically relevant model was a mean-derived correction of −7.5 HU that performed equal to all other mean corrections and regression corrections and was independent of sex or contrast timing phase. In addition, the −7.5 HU correction placed 40–45% of participants in the validation sample within 3 HU or less of baseline non-contrast and 63–70% of participants in the validation sample within 5 HU or less, similar to other corrections we tested. Other mean-derived and regression-derived correction tested similarly within the 3 HU contrast adjustment target, but were gender and phase specific, or required use of a regression equation. Therefore, we recommend the mean correction of −7.5 HU applied to either the AC or VC as the best method for estimating non-contrast density from a contrast scan independent of age, sex, or BMI. This simple correction is easily applicable for physicians and researchers, and helps harmonize contrast and non-contrast CT images for muscle density analysis.

As expected, we found significant muscle density differences between contrast and non-contrast images in our calibration sample, with contrast images having an 8.7 HU (18%) and a 6.4 HU (13%) increase for AC and VC, respectively. Our venous phase data, which was ensured to have consistent kV, differs slightly from previous research^15,16,19 by showing lower HU values than arterial phase data. Boutin et al. also found contrast increased paraspinal muscle density measurements, but by 15% in the arterial phase and by 25% in the venous phase.¹⁶ Using a fully-automated deep learning tool that includes the entire skeletal muscle cross-section at the L3 level,²⁰ Perez et al. reported a contrast-based increase of 18.8 HU in the parenchymal phase.¹⁰ This variation may be attributed to varying hospital protocols for contrast scans, including timing, dose, and scanner settings.

No matter the phase, contrast artificially increases muscle density, which raises issues when using CTs for diagnostic purposes. Recent sarcopenia diagnostic guidelines recommend the use of muscle quality and strength over muscle quantity.^12,13 Using CT muscle density as a surrogate of muscle quality is advantageous for several reasons: it is objective, independent of patient volition, and accessible to all patients who receive a CT scan regardless of their cognitive, mobility, or conscious status. Muscle density is also predictive of survival in many clinical populations, including heart failure,^21,22 cancer,^3,23,24 COVID-19,²⁵ and many others.^2,26–30 Before CT muscle density can be effectively utilized as a method of diagnosing muscle wasting diseases and predicting survival, cut-points for healthy muscle density must be established and be applicable to contrast and non-contrast exams. To this end, contrast and non-contrast data should not be used interchangeably without the application of a correction to account for artificial density inflation.¹¹ We found that the −7.5 HU correction worked well for both contrast phase image analyses and both sexes, suggesting that it may be broadly applicable to various hospital contrast protocols.

Our study has a few limitations. Our population consisted of patients weighing between 58.97 kg and 113.40 kg who underwent CT scans at 120kV on GE scanners with a weight-based contrast dose, which is not representative of all patients. In addition, because this study utilized bolus-tracked venous phase scans, the arterial time point was taken at the start of bolus tracking, 40 seconds after injection, which is slightly later than the typical arterial phase scan (up to 35 seconds after injection). Finally, our protocol used a variable contrast dose that increased linearly with patient weight, yet sites who use a fixed volume of contrast will observe varying CT number enhancement; higher enhancement for smaller weight patients and less enhancement for larger weight patients.¹⁸

Our simple correction of −7.5 HU applied to contrast scans may be useful for other patient populations and is clinically feasible for clinicians to implement. Future research should test this correction in populations with varying CT protocols, such as different kV, contrast dose parameters, various scanner makes and models, and different institutions. Additionally, it should be determined if our correction is applicable to other muscle measures, such as paraspinal or psoas whole-muscle HU, total axial skeletal muscle HU, and multi-slice or whole-body muscle HU. Moreover, virtual non-contrast (VNC) density measures may be calculated by digitally adjusting for iodine enhancement when dual-energy CT (DECT) is conducted,³¹ and future studies should determine how our correction performs against VNC. Furthermore, while we focused on contrast, other confounders exist, such as CT tube kilovoltage (kV), which can increase muscle density as much as 40% at lower kV compared to higher kV.³² These other factors may need to be incorporated into a correction model if variable levels of kV are used in the same sample population.