| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Experimental Studies |
1 From the Department of Radiology, New York University Medical Center, 560 First Ave, IRM 232, New York, NY 10016. Received December 19, 2005; revision requested February 9, 2006; revision received April 6; final version accepted May 8. Address correspondence to B.A.B. (e-mail: bernard.birnbaum{at}nyumc.org).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
Materials and Methods: A customized phantom was designed with tissue-equivalent materials to simulate contrast material–enhanced liver, spleen, pancreas, aorta, kidney, 0- and 50-HU cylindric renal cysts, muscle, and fat. The phantom was scanned with five multi–detector row CT scanners (LightSpeed QXi, GE Healthcare, Milwaukee, Wis; MX8000, Philips Medical Systems, Best, the Netherlands; and Volume Zoom, Sensation 16 and Sensation 64, Siemens Medical Solutions, Forchheim, Germany) on five separate occasions with 120 kVp, low and high tube current settings, 3.00–3.75-mm section thickness, 50% overlap, and standard and high-spatial-resolution kernels. Standardized regions of interest (ROIs) were used to obtain 3510 attenuation measurements. Attenuation dependence on scanner, kernel, and tube current was evaluated by using F tests derived with mixed-model regression. Within the mixed-model framework, the Tukey honestly significant difference procedure and a Bonferroni multiple comparison correction were used to assess differences among imaging regimens and tube current settings, respectively, in terms of tissue attenuation and ROI standard deviation.
Results: Tube current had no significant effect (P > .4) on observed tissue attenuation. Significant (P < .0001) differences were observed between imaging regimens with respect to mean attenuation for each tissue type. Convolution kernel modification had an inconsistent effect on tissue attenuation, depending on the scanner. All multi–detector row CT scanners displayed intrascanner variability in tissue attenuation (minimum range: 8.4 HU for fat tissue with the Sensation 16; maximum range: 63.4 HU for liver tissue with the Sensation 64). The scanners behaved differently at the lower range of the CT number scale, where 0-HU cyst attenuation ranged from –15.7 to 23.9 HU and one vendor's equipment showed significantly lower mean attenuation values.
Conclusion: CT attenuation values vary significantly between different manufacturers' multi–detector row CT scanners, among different generations of multi–detector row CT scanning equipment, and with individual combinations of scanner and convolution kernel.
Supplemental material: http://radiology.rsnajnls.org/cgi/content/full/242/1/109/DC1
© RSNA, 2007
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
Current methods for CT image reconstruction are not conducive to the calculation of absolute linear attenuation coefficients. The reconstructed pixel values reflect relative linear attenuation coefficients whose CT numbers can be compared by using a CT number scale in which –1000 represents the attenuation of air, 0 is the attenuation of water, and no upper limit exists (1). With the potential exceptions of values for fat, simple fluid, and potentially fresh hemorrhage, there are no typical CT attenuation values that permit the specific characterization of tissue type. Despite this central tenet, quantitative CT applications continue to evolve in which absolute CT numbers have been used for diagnostic purposes, including differentiation of benign and malignant disease processes (5–14).
The purpose of our study was to determine the dependence of absolute CT attenuation values on multi–detector row CT scanner type, convolution kernel, and tube current by using an anthropomorphic phantom.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
|
|
|
The principal phantom component was manufactured in 1999 with CT attenuation values that were calibrated by using a conventional CT scanner (CT/T 9800 HiLight scanner; GE Medical Systems, Milwaukee, Wis). The customized renal insert was manufactured in 2004 with CT attenuation values calibrated by using a helical CT scanner (High-Speed Advantage; GE Healthcare, Milwaukee, Wis). These calibrations were performed by Computerized Imaging Reference Systems with their own CT systems. The phantom core was financed by using prize monies awarded by the Society of Computed Body Tomography and Magnetic Resonance. The customized phantom insert was funded by Siemens Medical Solutions (Erlangen, Germany). All of the study authors had control of the study data and information submitted for publication.
Multi–Detector Row CT
The anthropomorphic phantom was imaged with five multi–detector row CT scanners on five separate occasions (25 total scanning sessions; temporal range between sessions, 6–36 days). The CT systems included three four-section multi–detector row CT scanners (Volume Zoom, Siemens, Forchheim, Germany [hereafter referred to as the Siemens 4 scanner]; LightSpeed QXi, GE Healthcare; and MX8000 Quad, Philips Medical Systems, Best, the Netherlands), a 16-section multi–detector row CT scanner (Sensation 16; Siemens) and a 64-section multi–detector row CT scanner (Sensation 64; Siemens). All five CT scanners were calibrated according to the manufacturer's specifications immediately before each phantom data acquisition with a complete calibration scan.
Image acquisition and reconstruction parameters for the individual CT scanners were as follows: Parameters for the Siemens 4 were 120 kVp; 50 and 200 mAs; 3.0-mm section thickness; 2.5-mm detector collimation; 1.25 beam pitch; 50% reconstruction interval; 33-cm field of view; 0.5-second gantry rotation time; and B40, B46, and B70 reconstruction kernels. Parameters for the LightSpeed QXi were 120 kVp, 50 and 200 mA (nominal tube current), 26.7 and 106.7 mAs (effective tube current), 3.75-mm section thickness, 2.5-mm detector collimation, 1.5 beam pitch, 50% reconstruction interval, 33-cm field of view, 0.8-second gantry rotation time, and standard and Bone Plus reconstruction kernels. Parameters for the MX8000 Quad were 120 kVp, 50 and 200 mAs, 3.0-mm section thickness, 1.0-mm detector collimation, 1.75 beam pitch, 50% reconstruction interval, 33-cm field of view, 0.75-second gantry rotation time, and B and EC reconstruction kernels. Parameters for the Sensation 16 were 120 kVp; 50 and 200 mAs; 3.0-mm section thickness; 0.75-mm detector collimation; 1.0 beam pitch; 50% reconstruction interval; 33-cm field of view; 0.5-second gantry rotation time; and B40, B46, and B70 reconstruction kernels. Finally, parameters for the Sensation 64 were 120 kVp; 50 and 200 mAs; 3.0-mm section thickness; 0.6-mm detector collimation; 1.4 beam pitch; 50% reconstruction interval; 33-cm field of view; 0.5-second gantry rotation time; and B40, B46, and B70 reconstruction kernels. Automatic dose modulation software was not used in this experiment.
A total of 130 phantom data sets were acquired. This number represents the product of scanning the phantom with 13 different imaging regimens (fixed combinations of multi–detector row CT scanner and convolution kernel) at two different tube current settings while evaluating each CT scanner five times.
Data Collection and Analysis
Absolute CT attenuation data were obtained for nine anatomic tissue types within the phantom (liver, spleen, pancreas, aorta, rectus abdominis muscle, body wall fat, kidney, and 0- and 50-HU cylindric renal cysts). The attenuation value of each anatomic structure was determined by drawing a circular region of interest (ROI) over each tissue type and recording the resultant attenuation measurement (in Hounsfield units) and ROI standard deviation displayed on the CT workstation. ROIs were drawn by one of two investigators (N.H. and J.L.). To minimize variation in individual performance, we held an initial training session to define specific fixed locations within each organ for ROI placement and to instruct each investigator in how to select ROIs of a standard size on the basis of the dimensions of the tissue of interest. ROI areas measured 0.21 cm2 (177 pixels), 0.43 cm2 (372 pixels), and 0.97 cm2 (837 pixels) for the 7-, 10-, and 15-mm cylindric cysts, respectively. ROI area measured 0.43 cm2 (372 pixels) for the parenchymal organs, aorta, muscle, and fat.
A total of 3510 CT attenuation measurements were acquired for this study. This was accomplished by generating 27 tissue-specific ROIs for each phantom data acquisition (n = 130). ROIs were placed over each anatomic structure of interest (n = 9) at three z-axis levels of the phantom that corresponded to the centers of the 7-, 10-, and 15-mm cylindric renal cysts. This ROI method enabled us to obtain three attenuation measurements for each tissue type during each phantom data acquisition (n = 130). This yielded a total of 390 attenuation measurements for each tissue type and a total of 3510 attenuation measurements for all anatomic structures studied. Partial volume effects were negated in the experiment by placing ROIs at the mid–z-axis level of each cyst cylinder and by using a 7.3-cm-deep slab phantom design (15).
Statistical Analysis
The CT attenuation values observed for each tissue type were summarized in terms of the mean, the standard deviation of the mean, and the median, minimum, maximum, and range of the values generated with each individual multi–detector row CT scanner over all relevant convolution kernels. Attenuation value variability was also assessed by analyzing the minimum, maximum, and range of attenuation values observed for each tissue type according to individual imaging regimen. F tests from mixed-model regression were used to evaluate the effect of imaging regimen and tube current setting on the attenuation values observed for the liver, spleen, pancreas, aorta, rectus abdominis muscle, body wall fat, kidney, and 0- and 50-HU renal cysts.
A separate analysis was conducted for each tissue type. In each case, the 390 observed tissue attenuation values (from three ROIs for each of 130 data acquisitions) constituted the dependent variable, while the model included imaging regimen and tube current setting as fixed classification factors and the term representing their interaction. The covariance structure was modeled by treating scanning session as a random classification factor, by using a compound symmetry structure for the covariance among observations from the three ROIs within the same data acquisition, and by allowing the error variance to differ across the levels of each fixed-effects factor. This implies that observations were correlated or independent when generated during the same or different scanning sessions, respectively, with the strength of correlation dependent on whether the observations were derived from the same data set. The results from the mixed-model regression analysis included a 95% prediction interval for the attenuation value of each individual tissue type as determined with each imaging regimen.
Mixed-model regression was also used to assess the effect of tube current setting and imaging regimen on the ROI standard deviations. For each tissue type, the mixed-model analysis followed the same outline described above, with the only exception being that the natural log of the ROI standard deviation was used as the dependent variable (distributional assumptions were best met after the log transformation).
Type 3 P values were used to assess the effect of each fixed-effect factor on the ROI standard deviation and attenuation value. Results were considered statistically significant at a family-wise 5% significance level. Specifically, the Tukey honestly significant difference (HSD) procedure was used to make pairwise comparisons among imaging regimens while maintaining the family-wise type I error rate for the set of comparisons at or below the 5% level. The effect of tube current setting was evaluated for each individual imaging regimen with a Bonferroni correction applied to control the family-wise type I error rate. That is, for each of the 13 imaging regimens, the effect of tube current setting was considered statistically significant only if associated with a type 3 P value of less than .05/13. All statistical computations were performed by using software (SAS for Windows, version 9.0, 2002; SAS Institute, Cary, NC).
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
|
|
Attenuation Value Variability
For the attenuation values observed for each phantom tissue type on each CT scanner over all relevant convolution kernels (Table 3), data revealed that substantial intra- and interscanner variability was present for all nine tissue types. When analyzed by individual CT scanner across all relevant kernels, the range (maximum minus minimum value) of the CT attenuation values for individual tissue types had a minimum of 8.4 HU for fat with the Siemens Sensation 16 (mean, –119.1 HU; range, –122.2 to –113.8 HU) and a maximum of 63.4 HU for the liver with the Siemens Sensation 64 (mean, 107.7 HU; range, 63.5–126.9 HU).
|
Intra- and Interscanner Variability
Significant differences between imaging regimens were noted in 100% (nine of nine) of the tissue types analyzed in this study (eg, the observed differences in mean attenuation when the 0-HU renal cyst was scanned across 13 imaging regimens were significant [Table 4]). When convolution kernels were compared between specific CT scanners, no significant differences (Tukey HSD–corrected P > .1) between kernels were identified for the Siemens 4, GE QXi, and Philips MX8000 scanners with respect to the mean attenuation observed for the 0-HU renal cyst. For the Siemens Sensation 16 scanner, the B46 kernel was associated with a significantly higher mean attenuation than the B70 kernel (Tukey HSD–corrected P = .0029). For the Siemens Sensation 64 scanner, the B40 kernel was associated with a significantly lower mean attenuation than either the B46 or the B70 kernel (Tukey HSD–corrected P < .001), while the latter two kernels were statistically indistinguishable (Tukey HSD–corrected P > .9). The results also demonstrate that the mean attenuation for the 0-HU renal cyst was significantly higher for the GE QXi and Philips MX8000 scanners than for each of the three Siemens scanners (Tukey HSD–corrected P < .001).
|
Convolution kernel modification had an inconsistent effect on tissue attenuation in the sense that the effect was found to depend on the individual CT scanner (Table 4; Tables E1–E8, http://radiology.rsnajnls.org/cgi/content/full/242/1/109/DC1). Varying the convolution kernel between the B40, B46, and B70 settings on the Siemens 4 scanner had no significant effect on the observed mean attenuation of any tissue type (Tukey HSD–corrected P > .1). In contrast, adjusting the convolution kernel between the B and EC settings on the Philips MX8000 scanner resulted in highly significant differences (Tukey HSD–corrected P < .001) in observed mean attenuation in 89% (eight of nine) of tissues studied, with the exception of the 0-HU renal cyst (Tukey HSD–corrected P = .99). Convolution kernel modification resulted in significant differences (Tukey HSD–corrected P < .05) in observed mean attenuation in 11% (one of nine) of tissue types (ie, only the 0-HU renal cyst) with the Siemens Sensation 16, in 44% (four of nine) of tissue types with the Siemens Sensation 64, and in 67% (six of nine) of tissue types with the GE QXi.
Scanner Linearity
Regarding CT scanner linearity (Fig 2), the data revealed that scanner performance closely tracked the reference attenuation values for all tissue types. The attenuation data obtained when scanning the 0-HU renal cyst tissue type with standard convolution kernels (Fig 3) demonstrated that the mean attenuation values for the 0-HU renal cyst with the Siemens scanners approximated 0 HU, while the mean attenuation values with the GE and Philips scanners approximated 9–10 HU. Although substantial overlap in absolute tissue attenuation values was noted, the observed mean attenuation values for the 0-HU renal cyst obtained with the GE and Philips scanners were significantly greater (Tukey HSD–corrected P < .001) than those obtained with the Siemens scanners (Table 4).
|
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
As expected, variation in tube current had a profound impact on image noise. As shown in our study, mean ROI standard deviation levels increased significantly for each and every tissue type as tube current decreased by a factor of four. When analyzed according to specific imaging regimen, the mean ROI standard deviations increased by an average of 95% ± 19. These results are congruent with the physics of CT, which dictate that image noise is inversely related to the square root of tube current.
Statistical analysis of our data led to an unexpected finding. At each tube current setting for each multi–detector row CT scanner other than the Philips MX8000, the mean ROI standard deviation was significantly lower with the standard kernel than with any high-spatial-resolution kernel used with the same scanner. For the Philips scanner, however, the B and EC kernels were statistically indistinguishable in this regard. The data indicate that image noise increased with the Siemens and GE scanners as progressively higher-resolution kernels were used at any given tube current setting. This effect was not observed with the Philips MX8000 scanner, for which image noise did not significantly increase when the higher-resolution EC kernel was tested. The results with the Philips scanner refute the popular notion that use of higher-resolution convolution kernels is always associated with increased image noise.
It should be clear from examination of the attenuation data we obtained that significant intra- and interscanner variability in absolute multi–detector row CT attenuation values may be observed and that convolution kernel modification may result in significant differences in tissue attenuation for a given multi–detector row CT scanner. Our results validate the findings of Levi et al (2), who noted the unreliability of CT numbers as absolute values. In our experiment, CT numbers varied considerably between different manufacturers' scanners, among different generations of scanning equipment, and with individual combinations of multi–detector row CT scanner and convolution kernel. Consequently, extreme caution should be exercised when one is attempting to use absolute CT numbers to characterize specific tissue types. Dedicated calibration phantoms have been successfully used to standardize pulmonary nodule, vertebral bone mineral, and coronary artery calcium quantitative analyses (16–19). The results of our study suggest that the use of CT calibration phantoms may be required for all quantitative CT applications to successfully adjust attenuation values and account for the known variability in CT numbers between different manufacturers' scanners, among different generations of scanning equipment, and within individual scanners.
The multi–detector row CT scanners evaluated in our study behaved differently at the lower range of the CT number scale. Despite substantial overlap in observed tissue attenuation, the mean attenuation values for the 0-HU renal cyst tissue type obtained with the GE and Philips scanners were significantly greater than those obtained with the Siemens scanners. This was true for all convolution kernels tested. Although all manufacturers strive to calibrate their scanners so that simple fluid corresponds to 0 HU, our data confirmed that the equipment we tested behaved differently in this regard and suggest that CT manufacturers likely use different methods for calibrating their multi–detector row CT scanners. When imaging regimens that utilized standard convolution kernels were compared in our experiment, the mean attenuation values for the 0-HU renal cyst approximated 0 HU with the Siemens scanners and 9–10 HU with the GE and Philips scanners. We believe that this finding has important clinical implications for CT densitometry applications that address the utility of absolute CT numbers in the range of 0–20 HU.
One such application is the use of absolute CT numbers to characterize cystic renal masses. A discriminatory threshold of 20–25 HU, derived from experience with conventional CT scanners, has historically been used to help differentiate simple renal cysts from complex cysts or solid lesions (6,8). However, if multi–detector row CT scanners are calibrated differently at the lower range of the CT number scale, as the data from our study suggest, the use of a single discriminatory threshold value or threshold range for all multi–detector row CT scanners appears unjustified. Instead, radiologists should adopt threshold criteria that are machine (imaging regimen) specific. On the basis of our study findings, we believe that lower discriminatory threshold criteria should be used for characterizing renal cysts with Siemens multi–detector row CT scanners than with GE or Philips multi–detector row CT equipment.
The second quantitative CT application that is affected by our study findings is adrenal CT dosimetry. Investigators have proposed that nonenhanced CT dosimetry can reliably enable differentiation of benign lipid-rich adrenal adenomas from lipid-poor adrenal adenomas and nonadenomatous adrenal lesions with 98% specificity if a 10-HU CT threshold criterion is used (11,12). In view of the substantial intra- and interscanner variability in multi–detector row CT attenuation values demonstrated in our study, as well as the observed differences in CT scanner performance in the range of 0–20 HU, we believe that it is highly unlikely that these published results can be directly extrapolated to all manufacturers' multi–detector row CT scanners, to varying generations of scanning equipment, or even to a given multi–detector row CT scanner operating with different convolution kernels. We propose that in practice, these quantitative CT applications should be deemphasized unless they are coupled with the use of calibration phantoms that incorporate standardization and calibration criteria that are specific for a particular imaging regimen.
Our phantom study had several limitations. Scan acquisition parameters for every scanner evaluated were not standardized owing to differences in multi–detector row CT design. Although this prevented us from always generating thin-section CT sections of the same thickness, we believe that we were able to avoid volume averaging effects in this experiment by using ROIs that were appropriately sized relative to the structure of interest, renal cyst cylinders instead of spheres, and a slab phantom design. Two investigators shared the responsibility for ROI placement and interpretation; however, we were able to minimize individual performance variation by training the investigators to use standardized ROI placement parameters. Finally, we did not evaluate every multi–detector row CT scanner in the marketplace, nor did we attempt to study every convolution kernel offered by each manufacturer. This was beyond the scope of our study. Despite this important limitation, however, we believe that the variability in multi–detector row CT scanner performance that we observed is likely applicable to all multi–detector row CT equipment.
Practical application: Results of our phantom study confirmed that absolute CT numbers may vary significantly between different manufacturers' multi–detector row CT scanners, among different generations of multi–detector row CT scanning equipment, and with individual combinations of multi–detector row CT scanner and convolution kernel. As a result, radiologists should exercise extreme caution before attempting to use absolute CT numbers for characterizing tissue type. Quantitative CT applications based on absolute CT attenuation values should be deemphasized unless they are coupled with the use of calibration phantoms that incorporate standardization and calibration criteria that are specific for a particular imaging regimen (multi–detector row CT scanner and convolution kernel combination). This is particularly true for quantitative applications in the range of 0–20 HU, because significant differences in multi–detector row CT performance may be observed at the lower range of the CT number scale.
|
ADVANCES IN KNOWLEDGE |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
|
FOOTNOTES |
|---|
Abbreviations: HSD = honestly significant difference • ROI = region of interest
2 Current address: Department of Radiology, Beth Israel Deaconess Medical Center, Boston, Mass 
3 Current address: Department of Radiologic Sciences, UCLA Medical Center, Los Angeles, Calif
See Materials and Methods for pertinent disclosures
Author contributions: Guarantor of integrity of entire study, B.A.B.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; literature research, B.A.B.; experimental studies, B.A.B., N.H., J.L.; statistical analysis, J.S.B.; and manuscript editing, all authors
|
References |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  L. M. Ho, E. K. Paulson, M. J. Brady, T. Z. Wong, and S. T. Schindera Lipid-Poor Adenomas on Unenhanced CT: Does Histogram Analysis Increase Sensitivity Compared with a Mean Attenuation Threshold? Am. J. Roentgenol., July 1, 2008; 191(1): 234 - 238. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. B. Reeder, F. Ranallo, and A. J. Taylor CT and MRI for Determining Hepatic Fat Content Am. J. Roentgenol., February 1, 2008; 190(2): W167 - W167. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
