Model-based Iterative Reconstruction Compared to Adaptive Statistical Iterative Reconstruction and Filtered Back-projection in CT of the Kidneys and the Adjacent Retroperitoneum Eric W. Olcott, MD, Lewis K. Shin, MD, Graham Sommer, MD, Ian Chan, MD, Jarrett Rosenberg, PhD, F. Lior Molvin, RT, F. Edward Boas, MD, PhD, Dominik Fleischmann, MD Rationale and Objectives: To prospectively evaluate the perceived image quality of model-based iterative reconstruction (MBIR) compared to adaptive statistical iterative reconstruction (ASIR) and filtered back-projection (FBP) in computed tomography (CT) of the kidneys and retroperitoneum. Materials and Methods: With investigational review board and Health Insurance Portability and Accountability Act compliance, 17 adults underwent 31 contrast-enhanced CT acquisitions at constant tube potential and current (range 30–300 mA). Each was reconstructed with MBIR, ASIR (50%), and FBP. Four reviewers scored each reconstruction’s perceived image quality overall and the perceived image quality of seven imaging features that were selected by the authors as being relevant to imaging in the region and pertinent to the evaluation of high-quality diagnostic CT. Results: MBIR perceived image quality scored superior to ASIR and FBP both overall (P < .001) and for observations of the retroperitoneal fascia (99.2%), corticomedullary differentiation (94.4%), renal hilar structures (96.8%), focal renal lesions (92.5%), and mitigation of streak artifact (100.0%; all, P < .001). MBIR achieved diagnostic overall perceived image quality with approximately half the radiation dose required by ASIR and FBP. The noise curve of MBIR was significantly lower and flatter (P < .001). Conclusions: Compared to ASIR and FBP, MBIR provides superior perceived image quality, both overall and for several specific imaging features, across a broad range of tube current levels, and requires approximately half the radiation dose to achieve diagnostic overall perceived image quality. Accordingly, MBIR should enable CT scanning with improved perceived image quality and/or reduced radiation exposure. Key Words: Computed tomography; image reconstruction; radiation protection; tomography scanners; x-ray computed; kidney. ªAUR, 2014
I
terative image reconstruction algorithms, such as the original algebraic reconstruction technique, (ART) were used in the early days of transmission computed tomography (CT) (1,2) but were quickly superseded by much faster analytical methods such as filtered back-projection (FBP) (3).
Acad Radiol 2014; 21:774–784 From the Department of Radiology, Stanford University School of Medicine, 300 Pasteur Drive, H1307, Stanford, CA 94305-5105 (E.W.O., L.K.S., G.S., I.C., J.R., F.L.M., F.E.B., D.F.) and Veterans Affairs Palo Alto Health Care System, Palo Alto, CA (E.W.O., L.K.S.). Received October 8, 2013; accepted February 10, 2014. Address correspondence to: E.W.O. e-mail: eolcott@ stanford.edu ªAUR, 2014 http://dx.doi.org/10.1016/j.acra.2014.02.012
774
Significant disadvantages exist with FBP, however, including its assumption that data are exact when in reality they are not, and the operation of the FBP filter which typically amplifies noise in the projection data. Iterative techniques, in contrast, can include models of features such as noise that improve the image during each iteration. This produces much better image quality than FBP does, when the signal-to-noise ratio is low, although this occurs at the expense of substantially increased computation time (4,5). Owing to substantial increases in computational power, iterative image reconstruction algorithms have recently re-emerged and are now available on commercial CT scanners manufactured by all major vendors (6). The recent literature appears to confirm the expected dose saving potential of
Academic Radiology, Vol 21, No 6, June 2014
modern implementations of iterative reconstruction algorithms (7–16); however, the relationship between radiation dose, noise, and image quality over a broad dose range, rather than at just a pair of dose levels, is not well established. The purpose of this study was to prospectively assess the relationship between radiation dose and perceived image quality for two commercially available iterative image reconstruction algorithms, model-based iterative reconstruction (MBIR) and adaptive statistical iterative reconstruction (ASIR), using comparisons of MBIR to ASIR and FBP. The term ‘‘perceived image quality’’ is used to indicate the reader-based nature of the image quality assessments, each conducted under defined experimental conditions. MBIR operates to mitigate the deleterious effects of both noise and geometric features of the scanner, whereas ASIR, the earlier of the two iterative algorithms, operates to mitigate noise (4,6–9,11–13,17). To perform this evaluation in a welldefined anatomic region with tissues of substantially varying density, we chose to focus specifically on the kidneys and the adjacent retroperitoneum. Seven anatomic features were chosen by the authors for specific assessment because they relate to essential anatomy and imaging challenges in this region and, accordingly, were felt pertinent to the evaluation of high-quality diagnostic CT. Each such feature was additionally felt to be readily recognizable by radiologists and amenable to evaluation using straightforward criteria.
MATERIALS AND METHODS Study Population
Patients scheduled for outpatient clinical abdominal contrast-enhanced CT scans were consecutively offered the opportunity to participate until 18 individuals, the total for which the study was designed, had agreed to participate and provided informed consent in compliance with our investigational review board (IRB) and the Health Insurance Portability and Accountability Act (HIPAA). No attempts at patient selection were made. Because the digital data for one patient became unexpectedly unusable owing to an electronic corruption event, analysis was performed on the remaining 17 patients. The indications for CT scanning among the 17 patients included abdominal pain in one and evaluation for known or possible oncologic disease in 16. The 17 patients included 7 female adults (age range 29–60 years, mean 46.9 years) and 10 male adults (age range 33–72 years, mean 49.4 years). Scanning Technique and Image Reconstruction
Clinical CT scans of the abdomen and pelvis were acquired with a 64-detector row scanner (General Electric Discovery HD750; General Electric Healthcare, Waukesha, WI). Technical factors for all scans included a 64 0.625 mm detector configuration, pitch = 1.375:1, rotation time = 0.8 seconds, 32-cm field of view, and large scan field of view. For each
MBIR OF KIDNEYS AND RETROPERITONEUM
scan, iohexol (Omnipaque 350; GE Healthcare, Milwaukee, WI) was administered at 1.6 mL/kg body weight and injected intravenously within a period of 40 seconds. At our institution, delayed excretory-phase images specifically through the kidneys are routinely acquired at 2.5–5.0 minutes after contrast injection for abdominal CT, in the excretory phase of renal contrast enhancement, and these excretory-phase images form the basis for the present study. We specifically elected to focus on excretory phase imaging to avoid disrupting the timing of earlier phases of clinical imaging and because it provides more constant levels of enhancement. Thirty-one such 2.5- to 5.0minute delayed scans (or ‘‘acquisitions’’) were performed among the 17 patients, all obtained at 120 kV with constant tube current, and to remove variability in tube current, no tube current modulation. Each patient experienced a total volume computed tomography dose index (CTDIvol) level of 14.0 mGy delivered as either one delayed scan performed at 300 mA or two delayed scans whose CTDIvol values summed to 14.0 mGy and whose tube current values summed to 300 mA (eg, 30 and 270 mA; Table 1). Acquisitions were obtained with tube current values stepped from 30 to 300 mA among the group of patients but held constant for each individual patient, in the order indicated in Table 1, with most steps equivalent to 10% increments of the 14.0 mGy summed exposure value. For the purposes of this study and separately from clinical interpretation, the raw data were anonymized and transferred to the scanner manufacturer (General Electric Healthcare, Waukesha, WI) for reconstruction as our institution had no on-site reconstruction capability. Images were reconstructed with FBP, ASIR (50%), and MBIR and subsequently returned for review in DICOM format. Image Review
Four radiologists independently evaluated the image sets generated with each of the three algorithms from each of the 31 acquisitions. The reviewers’ experience in diagnostic radiology ranged from 5 to over 35 years. The three reconstructed image sets (or ‘‘reconstructions’’) from each acquisition were displayed on a workstation (AquariusNet; TeraRecon Inc., San Mateo, CA) and evaluated under typical clinical viewing conditions in a darkened room with clinical monitors and interactive mouse and keyboard controls. The readers began with window/level settings of 400/40 Hounsfield units (HU) for soft tissue structures and 2000/300 HU for the intrarenal collecting systems and then altered these settings as desired while scrolling, magnifying, and panning at will. While neither milliampere values nor reconstruction algorithms appeared on the images, the unique image texture identifying MBIR reconstructions was readily apparent. The reviewers were not informed regarding the disease status of any patient, abnormalities that were or were not present, or the size or location of any abnormalities. Individual imaging features. For each acquisition, the reviewers independently assessed the perceived image quality of MBIR 775
OLCOTT ET AL
Academic Radiology, Vol 21, No 6, June 2014
TABLE 1. Scanning Parameters and Relative Doses in 17 Patients First CT Acquisition Number of Patients
Second CT Acquisition
Total Scans
Tube Current (mA)
Individual CTDIvol (mGy)
Fraction of Total CTDIvol (%)
Tube Current (mA)
Individual CTDIvol (mGy)
Fraction of Total CTDIvol (%)
3 6 6 4* 6 6
300 270 255 225 195 165
14 12.6 11.9 10.5 9.1 7.7
100 90 85 75 65 55
— 30 45 75 105 135
— 1.4 2.1 3.5 4.9 6.3
— 10 15 25 35 45
3 3 3 2* 3 3
ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered back-projection; MBIR, model-based iterative reconstruction; CTDIvol, volume computed tomography dose index. *Two scans from one of 18 original patients were lost to corruption of data, leaving reduced numbers as indicated.
versus ASIR and MBIR versus FBP in side-by-side fashion with respect to seven imaging features including the retroperitoneal fascia, corticomedullary differentiation, hilar structures, lesions (if present), intrarenal collecting system, calculi (if present), and the mitigation of streak artifacts. The first six of these points were evaluated on the basis of edge definition and homogeneity; the last, mitigation of streak artifact, was evaluated within the retroperitoneal fat. The perceived image quality of each imaging feature depicted by ASIR and FBP in each acquisition was scored as inferior, equivalent or superior to the quality depicted by MBIR. Each reviewer’s assessment of a given imaging feature in the reconstructions from one acquisition was counted as one observation. Overall perceived image quality. The four readers also independently evaluated the overall perceived image quality of the kidneys and adjacent retroperitoneum for each of the three reconstructions for each acquisition, using a four-point scale: 0 = nondiagnostic, 1 = minimally diagnostic, 2 = diagnostic, and 3 = superior. ‘‘Diagnostic’’ quality was defined as equivalent to that of well-performed ASIR scans that we see in our day-to-day clinical radiology practice. Each reconstruction was viewed separately under typical clinical viewing conditions as was done during review of individual imaging features. Randomization was not undertaken because MBIR reconstructions were readily recognizable by their unique image texture. However, each reconstruction was evaluated independently of all others, and the reviewers were blinded to milliampere values. Image noise. Image noise in each reconstruction from a given acquisition was measured within a circular region of interest deposited over the abdominal aorta and computed in the typical fashion as the standard deviation of attenuation values (HU) (18).
hypothesis that MBIR would be scored as superior with a probability of 0.50 owing to chance alone. To be able to consider the perceived image quality of individual features in a clinically familiar context, the corresponding data were also stratified into ranges defined by tube current and termed ‘‘standard dose’’ (255–300 mA), ‘‘low dose’’ (135–225 mA), and ‘‘very low dose’’ (30–105 mA). Overall perceived image quality. To evaluate the relationship between milliamperage and mean overall perceived image quality score as assessed within the kidneys and the adjacent retroperitoneum, nonlinear least-squares regression analysis was performed with curve-fitting, adjusting for clustering within patients. (Fig 5). Significance tests for differences between curve parameter values for the different reconstructions were done by the seemingly unrelated regression (SUR) method (19). Image noise. Nonlinear least-squares analysis with curvefitting and the SUR method were also used to evaluate the relationship between image noise and tube current. Interobserver agreement was evaluated with the kappa statistic. All statistical analyses were done using Stata Release 12.1 (StataCorp LP, College Station, TX). Investigational Approval
This investigation was conducted with approval of the IRB issued by our Panel on Human Subjects in Medical Research on May 31, 2011 and proceeded in compliance with the HIPAA and with informed consent.
RESULTS Individual Imaging Features
Statistical Analysis
Individual imaging features. For the evaluation of individual imaging features, the exact binomial test was applied to the proportion of observations in which MBIR was rated superior to both ASIR and FBP, testing against the null 776
The MBIR algorithm was rated superior to both ASIR and FBP with respect to most of the individual imaging features, when considered in total and when stratified into dose ranges of ‘‘standard dose’’ (255–300 mA), ‘‘low dose’’ (135–225 mA), and ‘‘very low dose’’ (30–105 mA; Table 2, Figs 1–4).
Academic Radiology, Vol 21, No 6, June 2014
MBIR OF KIDNEYS AND RETROPERITONEUM
Figure 1. Axial CT images acquired at 270 mA (a–c) and 30 mA (d–f) and reconstructed with MBIR (a, d), ASIR (b, e), and FBP (c, f). Perceived image quality of the retroperitoneal fascia (white arrows) and corticomedullary differentiation (black arrows) was superior with MBIR at both milliampere levels. Streak artifact is absent in the MBIR images but present in the ASIR and FBP images. Nodular appearance of the lower dose MBIR image (d) compared to the higher dose image (a) is noted. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered back-projection; MBIR, model-based iterative reconstruction.
Demographic data corresponding to stratification are presented in Table 3. In no observation of any individual feature was MBIR rated inferior to either ASIR or FBP; accordingly, observations in which MBIR was not rated superior are observations in which MBIR was equivalent to one or both of the other two algorithms. Many but not all acquisitions exhibited focal renal lesions, and only a few revealed calculi (Table 2). All focal renal lesions were hypoattenuating relative to the enhancing renal parenchyma, cortical and/or medullary in location, round to elliptical in shape, and 0.5–2.0 cm in greatest dimension. Interobserver agreement was high for mitigation of streak artifact (all readers scored MBIR superior to ASIR and FBP for all observations) and for perceived image quality of the perinephric fascia, hilar structures, and corticomedullary differentiation (three readers scored MBIR as superior to both ASIR and FBP for all observations of these features and the fourth reader did so for all but one of 31, four of 31, and seven of 31 observations of the three features, respectively). Interobserver agreement was also high for the perceived image quality of lesions, with MBIR being rated superior to ASIR in at least 90% of observations by each reader and superior to FBP in at least 95% of observations by each reader.
In contrast, MBIR was significantly not superior with respect to the perceived image quality of the intrarenal collecting system (Table 2), and the corresponding interobserver agreement was only moderate, with k = 0.75 (95% confidence interval [CI] = 0.48–0.92) for MBIR versus ASIR and k = 0.64 (95% CI = 0.32–0.83) for MBIR versus FBP. Of note, ‘‘significantly not superior’’ is not equivalent to ‘‘significantly inferior’’ and, rather, indicates that MBIR was most frequently scored as equivalent to ASIR and/or FBP in observations of this imaging feature. Overall Perceived Image Quality
Nonlinear least-squares regression analysis was performed to evaluate the relationship between milliamperage and mean overall perceived image quality score as assessed within the kidneys and the adjacent retroperitoneum (Fig 5). The data were well fit by a Gompertz function of the form. Image quality ¼ b1 exp ð exp ½ b2 fmA b3gÞ where b1, b2, and b3 are parameters with estimated values and 95% CIs as displayed in Table 4. Parameters b1 and b3 were 777
OLCOTT ET AL
Academic Radiology, Vol 21, No 6, June 2014
Figure 2. Axial CT images acquired at 255 mA (a–c) and 45 mA (d–f) and reconstructed with MBIR (a, d), ASIR (b, e), and FBP (c, f). Perceived image quality of hilar structures (arrowheads) and focal water-density lesion (arrows) was superior with MBIR at both milliampere levels. Streak artifact is absent in the MBIR images but present in the ASIR and FBP images. Nodular appearance of the lower dose MBIR image (d) compared to the higher dose image (a) is noted. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered backprojection; MBIR, model-based iterative reconstruction.
significantly different for MBIR compared to those for ASIR and FBP (both with P < .001), whereas parameter b2 did not differ significantly (P = .132 and .650, respectively). The MBIR algorithm required less milliamperage to reach any given overall perceived image quality level compared to ASIR and FBP and, similarly, achieved a higher level of overall perceived image quality across the range of tube current levels (Fig 5). The MBIR curve intersected the overall perceived image quality levels of 2 (diagnostic) and 1 (minimally diagnostic) at approximately 105 and 25 mA (by extrapolation), respectively, whereas the ASIR curve did the same at approximately 235 and 100 mA, respectively. The FBP curve behaved similarly to the ASIR curve but followed a somewhat lower course. Thus, MBIR required approximately 130 fewer milliamperes or 55.3% (1 105/235) less radiation dose than ASIR to produce an overall perceived image quality score of 2 (diagnostic) and approximately 75 fewer milliamperes or 75.0% (1 25/100) less dose than ASIR and FBP to achieve an overall perceived image quality score of 1 (minimally diagnostic). Moreover, at all milliampere levels the MBIR curve consistently exceeded the ASIR and FBP curves by 0.8–0.9 points in overall perceived image quality, and only 778
the MBIR curve approached or met the overall perceived image quality score of 3 (superior) at any milliampere level. With respect to overall perceived image quality scores, pairwise agreement among the four readers was moderate for all pairings except those involving reader 3 whose associations were less strong, as shown by the linearly weighted kappa statistic (Table 5). Table 6 indicates the minimum tube current levels at which all four individual readers reported overall perceived image quality scores of at least 1, at least 2, and at least 3 for ASIR, FBP, and MBIR. Of particular note, MBIR and ASIR required 135 and 270 mA, respectively, for scores of at least 2 (diagnostic). The values for FBP were equivalent to those of ASIR. All reviewers described the perceived image quality in lower-milliamperage MBIR images as ‘‘blotchy’’ or ‘‘nodular’’ compared to higher-milliamperage MBIR images whereas lower-milliamperage ASIR and FBP images were described as having familiar ‘‘point-like’’ patterns of noise compared to corresponding higher-milliamperage images (Figs 1,2). Image Noise
Nonlinear least-squares regression analysis was performed to evaluate the relationship between milliampere and mean
Academic Radiology, Vol 21, No 6, June 2014
MBIR OF KIDNEYS AND RETROPERITONEUM
Figure 3. Axial CT images acquired at 225 mA (a–c) and 75 mA (d–f), reconstructed with MBIR (a, d), ASIR (b, e), and FBP (c, f). Perceived image quality of a small urinary calculus (arrows) was equivalent among MBIR, ASIR, and FBP at the higher milliampere level and superior for MBIR at the lower milliampere level. Streak artifact is absent in the MBIR images but present in the ASIR and FBP images. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered back-projection; MBIR, model-based iterative reconstruction.
image noise as assessed within the aorta (Fig 6, Table 7). These data fit a power function of the form Noise ¼ b0 ðmAÞ ^b1 where b0 and b1 are parameters whose estimated values and 95% CIs are shown in Table 7. Both parameters b0 and b1 differed significantly between MBIR and ASIR and between MBIR and FBP (b0: P = .011 and .006, respectively; b1: P < .001 in both cases). Image noise for ASIR and FBP rose with decreasing milliamperes at a rate proportional to nearly the traditionally expected rate of (mA)0.5 (9,12) as evidenced by their b1 values of 0.53 and 0.51, respectively. In contrast, noise for MBIR rose with decreasing milliamperes at a much slower rate, with b1 = 0.22.
DISCUSSION This prospective investigation supports earlier reports that modern iterative image reconstruction algorithms such as MBIR should provide improved CT image quality when compared to images reconstructed with earlier algorithms
such as ASIR or FBP at given radiation dose levels (7–15). Similarly, our results indicate that algorithms such as MBIR produce a given level of perceived image quality with lower radiation dose compared to algorithms such as ASIR and FBP. Our conclusions are based on examinations of the kidneys and the adjacent retroperitoneum as visualized in the renal excretory phase of contrast enhancement. Our study design included comparisons of MBIR to ASIR and FBP along a stepwise gradation of dose levels, from tube current values of 30–300 mA (corresponding CTDIvol 1.4–14 mGy), with evaluation of overall perceived image quality in a stepwise fashion from a score of ‘‘nondiagnostic’’ through a score of ‘‘superior.’’ This design permits a more informative evaluation than one that just compares the clinical adequacy of image quality of one technology operating at a ‘‘typical’’ clinical dose to that of a newer technology operating at a lower dose, with no gradations of dose in between, such that apparent ‘‘dose saving’’ may result simply from a ‘‘typical’’ clinical dose that is higher than necessary to provide adequate image quality. Thus, our data permit a finely scaled comparison of the perceived image quality of MBIR compared to that of ASIR and FBP, as a function of dose. As shown by nonlinear least-squares regression, the MBIR algorithm produced 779
OLCOTT ET AL
Academic Radiology, Vol 21, No 6, June 2014
Figure 4. Axial CT images acquired at 270 mA (a–c) and 30 mA (d–f), reconstructed with MBIR (a, d), ASIR (b, e), and FBP (c, f) and displayed on wide window settings. Perceived image quality of the contrast medium–opacified intrarenal collecting system (arrows) was equivalent among MBIR, ASIR, and FBP at both milliampere levels. Streak artifact is absent in the MBIR images but present in the ASIR and FBP images. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered back-projection; MBIR, model-based iterative reconstruction.
Figure 5. Graph showing nonlinear least-squares regression relating tube current (mA) and mean perceived image quality scores for reconstructions performed with MBIR, ASIR, and FBP. Curves describe a Gompertz function of the form ‘‘Image quality = b1 exp (exp [b2 {dose b3}])’’. Vertical bars correspond to + and one standard error from the mean image quality score at each milliampere value. ASIR, adaptive statistical iterative reconstruction; FBP, filtered back-projection; MBIR, model-based iterative reconstruction.
780
significantly better overall perceived image quality than ASIR and FBP did at all dose levels (Fig 5, Table 4). To produce scans of ‘‘diagnostic’’ overall perceived image quality, MBIR required approximately 105 mA or approximately 130 fewer mA than ASIR required, which corresponds to 55.3% (1 105/235) less dose than ASIR required. Similarly, although an observation that is arguably less rigorous in nature, the minimum milliampere levels at which all four individual readers reported overall perceived image quality scores of at least ‘‘diagnostic’’ were 135 mA for MBIR and 270 mA for both ASIR and FBP (Table 6). These findings suggest that a ‘‘floor’’ or minimum dose necessary to produce scans with diagnostic quality in our study exists near 105–135 mA for MBIR, compared to 235–270 mA for ASIR and FBP in our study, representing a dose saving of 50% for MBIR compared to ASIR and FBP. This feature of MBIR may be useful in mitigating medical radiation exposure, a topic of considerable importance (20–26). In addition, MBIR produced superior depiction of several individual imaging features compared to ASIR and FBP and did so significantly when tested against the null hypothesis that MBIR would be scored as superior in 50% of observations
Academic Radiology, Vol 21, No 6, June 2014
MBIR OF KIDNEYS AND RETROPERITONEUM
TABLE 2. Observations of Individual Imaging Features in Which MBIR was Rated Superior to Both ASIR and FBP
Imaging Feature
Percentage of Observations in Which MBIR was Rated Superior
Very Low Dose 30–105 mA 1.4–4.9 mGy
Low Dose 135–225 mA 6.3–10.5 mGy
Standard Dose 255–300 mA 11.9–14.0 mGy
99.2% (123/124) 94.4% (117/124) 96.8% (120/124) 92.5% (74/80)
100.0% (44/44) 97.7% (43/44) 100.0% (44/44) 100.0% (32/32)
97.7% (43/44) 93.2% (41/44) 95.5% (42/44) 91.7% (22/24)
100.0% (124/124) 23.4% (29/124)
100.0% (44/44) 56.8% (25/44) P = .451 100.0% (4/4) P = .125
100.0% (44/44) 6.8% (3/44)
100.0% (36/36) 91.7% (33/36) 94.4% (34/36) 83.3% (20/24) P = .002 100.0% (36/36) 2.8% (1/36)
Retroperitoneal fascia Corticomedullary differentiation Hilar structures Lesions Mitigation of streak artifact Intrarenal collecting system
Percentage of Observations in Which MBIR was Rated Superior, Stratified by Tube Current (mA) and CTDIvol (mGy)
Calculi
6.3% (5/8) P = .727
25.0% (1/4) P = .625
—
ASIR, adaptive statistical iterative reconstruction; FBP, filtered back-projection; MBIR, model-based iterative reconstruction; CTDIvol, volume computed tomography dose index. Numerators indicate observations in which MBIR was rated superior to both ASIR and FBP; denominators indicate the total number of observations. Significance is assessed with the exact binomial test. All P values
I
terative image reconstruction algorithms, such as the original algebraic reconstruction technique, (ART) were used in the early days of transmission computed tomography (CT) (1,2) but were quickly superseded by much faster analytical methods such as filtered back-projection (FBP) (3).
Acad Radiol 2014; 21:774–784 From the Department of Radiology, Stanford University School of Medicine, 300 Pasteur Drive, H1307, Stanford, CA 94305-5105 (E.W.O., L.K.S., G.S., I.C., J.R., F.L.M., F.E.B., D.F.) and Veterans Affairs Palo Alto Health Care System, Palo Alto, CA (E.W.O., L.K.S.). Received October 8, 2013; accepted February 10, 2014. Address correspondence to: E.W.O. e-mail: eolcott@ stanford.edu ªAUR, 2014 http://dx.doi.org/10.1016/j.acra.2014.02.012
774
Significant disadvantages exist with FBP, however, including its assumption that data are exact when in reality they are not, and the operation of the FBP filter which typically amplifies noise in the projection data. Iterative techniques, in contrast, can include models of features such as noise that improve the image during each iteration. This produces much better image quality than FBP does, when the signal-to-noise ratio is low, although this occurs at the expense of substantially increased computation time (4,5). Owing to substantial increases in computational power, iterative image reconstruction algorithms have recently re-emerged and are now available on commercial CT scanners manufactured by all major vendors (6). The recent literature appears to confirm the expected dose saving potential of
Academic Radiology, Vol 21, No 6, June 2014
modern implementations of iterative reconstruction algorithms (7–16); however, the relationship between radiation dose, noise, and image quality over a broad dose range, rather than at just a pair of dose levels, is not well established. The purpose of this study was to prospectively assess the relationship between radiation dose and perceived image quality for two commercially available iterative image reconstruction algorithms, model-based iterative reconstruction (MBIR) and adaptive statistical iterative reconstruction (ASIR), using comparisons of MBIR to ASIR and FBP. The term ‘‘perceived image quality’’ is used to indicate the reader-based nature of the image quality assessments, each conducted under defined experimental conditions. MBIR operates to mitigate the deleterious effects of both noise and geometric features of the scanner, whereas ASIR, the earlier of the two iterative algorithms, operates to mitigate noise (4,6–9,11–13,17). To perform this evaluation in a welldefined anatomic region with tissues of substantially varying density, we chose to focus specifically on the kidneys and the adjacent retroperitoneum. Seven anatomic features were chosen by the authors for specific assessment because they relate to essential anatomy and imaging challenges in this region and, accordingly, were felt pertinent to the evaluation of high-quality diagnostic CT. Each such feature was additionally felt to be readily recognizable by radiologists and amenable to evaluation using straightforward criteria.
MATERIALS AND METHODS Study Population
Patients scheduled for outpatient clinical abdominal contrast-enhanced CT scans were consecutively offered the opportunity to participate until 18 individuals, the total for which the study was designed, had agreed to participate and provided informed consent in compliance with our investigational review board (IRB) and the Health Insurance Portability and Accountability Act (HIPAA). No attempts at patient selection were made. Because the digital data for one patient became unexpectedly unusable owing to an electronic corruption event, analysis was performed on the remaining 17 patients. The indications for CT scanning among the 17 patients included abdominal pain in one and evaluation for known or possible oncologic disease in 16. The 17 patients included 7 female adults (age range 29–60 years, mean 46.9 years) and 10 male adults (age range 33–72 years, mean 49.4 years). Scanning Technique and Image Reconstruction
Clinical CT scans of the abdomen and pelvis were acquired with a 64-detector row scanner (General Electric Discovery HD750; General Electric Healthcare, Waukesha, WI). Technical factors for all scans included a 64 0.625 mm detector configuration, pitch = 1.375:1, rotation time = 0.8 seconds, 32-cm field of view, and large scan field of view. For each
MBIR OF KIDNEYS AND RETROPERITONEUM
scan, iohexol (Omnipaque 350; GE Healthcare, Milwaukee, WI) was administered at 1.6 mL/kg body weight and injected intravenously within a period of 40 seconds. At our institution, delayed excretory-phase images specifically through the kidneys are routinely acquired at 2.5–5.0 minutes after contrast injection for abdominal CT, in the excretory phase of renal contrast enhancement, and these excretory-phase images form the basis for the present study. We specifically elected to focus on excretory phase imaging to avoid disrupting the timing of earlier phases of clinical imaging and because it provides more constant levels of enhancement. Thirty-one such 2.5- to 5.0minute delayed scans (or ‘‘acquisitions’’) were performed among the 17 patients, all obtained at 120 kV with constant tube current, and to remove variability in tube current, no tube current modulation. Each patient experienced a total volume computed tomography dose index (CTDIvol) level of 14.0 mGy delivered as either one delayed scan performed at 300 mA or two delayed scans whose CTDIvol values summed to 14.0 mGy and whose tube current values summed to 300 mA (eg, 30 and 270 mA; Table 1). Acquisitions were obtained with tube current values stepped from 30 to 300 mA among the group of patients but held constant for each individual patient, in the order indicated in Table 1, with most steps equivalent to 10% increments of the 14.0 mGy summed exposure value. For the purposes of this study and separately from clinical interpretation, the raw data were anonymized and transferred to the scanner manufacturer (General Electric Healthcare, Waukesha, WI) for reconstruction as our institution had no on-site reconstruction capability. Images were reconstructed with FBP, ASIR (50%), and MBIR and subsequently returned for review in DICOM format. Image Review
Four radiologists independently evaluated the image sets generated with each of the three algorithms from each of the 31 acquisitions. The reviewers’ experience in diagnostic radiology ranged from 5 to over 35 years. The three reconstructed image sets (or ‘‘reconstructions’’) from each acquisition were displayed on a workstation (AquariusNet; TeraRecon Inc., San Mateo, CA) and evaluated under typical clinical viewing conditions in a darkened room with clinical monitors and interactive mouse and keyboard controls. The readers began with window/level settings of 400/40 Hounsfield units (HU) for soft tissue structures and 2000/300 HU for the intrarenal collecting systems and then altered these settings as desired while scrolling, magnifying, and panning at will. While neither milliampere values nor reconstruction algorithms appeared on the images, the unique image texture identifying MBIR reconstructions was readily apparent. The reviewers were not informed regarding the disease status of any patient, abnormalities that were or were not present, or the size or location of any abnormalities. Individual imaging features. For each acquisition, the reviewers independently assessed the perceived image quality of MBIR 775
OLCOTT ET AL
Academic Radiology, Vol 21, No 6, June 2014
TABLE 1. Scanning Parameters and Relative Doses in 17 Patients First CT Acquisition Number of Patients
Second CT Acquisition
Total Scans
Tube Current (mA)
Individual CTDIvol (mGy)
Fraction of Total CTDIvol (%)
Tube Current (mA)
Individual CTDIvol (mGy)
Fraction of Total CTDIvol (%)
3 6 6 4* 6 6
300 270 255 225 195 165
14 12.6 11.9 10.5 9.1 7.7
100 90 85 75 65 55
— 30 45 75 105 135
— 1.4 2.1 3.5 4.9 6.3
— 10 15 25 35 45
3 3 3 2* 3 3
ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered back-projection; MBIR, model-based iterative reconstruction; CTDIvol, volume computed tomography dose index. *Two scans from one of 18 original patients were lost to corruption of data, leaving reduced numbers as indicated.
versus ASIR and MBIR versus FBP in side-by-side fashion with respect to seven imaging features including the retroperitoneal fascia, corticomedullary differentiation, hilar structures, lesions (if present), intrarenal collecting system, calculi (if present), and the mitigation of streak artifacts. The first six of these points were evaluated on the basis of edge definition and homogeneity; the last, mitigation of streak artifact, was evaluated within the retroperitoneal fat. The perceived image quality of each imaging feature depicted by ASIR and FBP in each acquisition was scored as inferior, equivalent or superior to the quality depicted by MBIR. Each reviewer’s assessment of a given imaging feature in the reconstructions from one acquisition was counted as one observation. Overall perceived image quality. The four readers also independently evaluated the overall perceived image quality of the kidneys and adjacent retroperitoneum for each of the three reconstructions for each acquisition, using a four-point scale: 0 = nondiagnostic, 1 = minimally diagnostic, 2 = diagnostic, and 3 = superior. ‘‘Diagnostic’’ quality was defined as equivalent to that of well-performed ASIR scans that we see in our day-to-day clinical radiology practice. Each reconstruction was viewed separately under typical clinical viewing conditions as was done during review of individual imaging features. Randomization was not undertaken because MBIR reconstructions were readily recognizable by their unique image texture. However, each reconstruction was evaluated independently of all others, and the reviewers were blinded to milliampere values. Image noise. Image noise in each reconstruction from a given acquisition was measured within a circular region of interest deposited over the abdominal aorta and computed in the typical fashion as the standard deviation of attenuation values (HU) (18).
hypothesis that MBIR would be scored as superior with a probability of 0.50 owing to chance alone. To be able to consider the perceived image quality of individual features in a clinically familiar context, the corresponding data were also stratified into ranges defined by tube current and termed ‘‘standard dose’’ (255–300 mA), ‘‘low dose’’ (135–225 mA), and ‘‘very low dose’’ (30–105 mA). Overall perceived image quality. To evaluate the relationship between milliamperage and mean overall perceived image quality score as assessed within the kidneys and the adjacent retroperitoneum, nonlinear least-squares regression analysis was performed with curve-fitting, adjusting for clustering within patients. (Fig 5). Significance tests for differences between curve parameter values for the different reconstructions were done by the seemingly unrelated regression (SUR) method (19). Image noise. Nonlinear least-squares analysis with curvefitting and the SUR method were also used to evaluate the relationship between image noise and tube current. Interobserver agreement was evaluated with the kappa statistic. All statistical analyses were done using Stata Release 12.1 (StataCorp LP, College Station, TX). Investigational Approval
This investigation was conducted with approval of the IRB issued by our Panel on Human Subjects in Medical Research on May 31, 2011 and proceeded in compliance with the HIPAA and with informed consent.
RESULTS Individual Imaging Features
Statistical Analysis
Individual imaging features. For the evaluation of individual imaging features, the exact binomial test was applied to the proportion of observations in which MBIR was rated superior to both ASIR and FBP, testing against the null 776
The MBIR algorithm was rated superior to both ASIR and FBP with respect to most of the individual imaging features, when considered in total and when stratified into dose ranges of ‘‘standard dose’’ (255–300 mA), ‘‘low dose’’ (135–225 mA), and ‘‘very low dose’’ (30–105 mA; Table 2, Figs 1–4).
Academic Radiology, Vol 21, No 6, June 2014
MBIR OF KIDNEYS AND RETROPERITONEUM
Figure 1. Axial CT images acquired at 270 mA (a–c) and 30 mA (d–f) and reconstructed with MBIR (a, d), ASIR (b, e), and FBP (c, f). Perceived image quality of the retroperitoneal fascia (white arrows) and corticomedullary differentiation (black arrows) was superior with MBIR at both milliampere levels. Streak artifact is absent in the MBIR images but present in the ASIR and FBP images. Nodular appearance of the lower dose MBIR image (d) compared to the higher dose image (a) is noted. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered back-projection; MBIR, model-based iterative reconstruction.
Demographic data corresponding to stratification are presented in Table 3. In no observation of any individual feature was MBIR rated inferior to either ASIR or FBP; accordingly, observations in which MBIR was not rated superior are observations in which MBIR was equivalent to one or both of the other two algorithms. Many but not all acquisitions exhibited focal renal lesions, and only a few revealed calculi (Table 2). All focal renal lesions were hypoattenuating relative to the enhancing renal parenchyma, cortical and/or medullary in location, round to elliptical in shape, and 0.5–2.0 cm in greatest dimension. Interobserver agreement was high for mitigation of streak artifact (all readers scored MBIR superior to ASIR and FBP for all observations) and for perceived image quality of the perinephric fascia, hilar structures, and corticomedullary differentiation (three readers scored MBIR as superior to both ASIR and FBP for all observations of these features and the fourth reader did so for all but one of 31, four of 31, and seven of 31 observations of the three features, respectively). Interobserver agreement was also high for the perceived image quality of lesions, with MBIR being rated superior to ASIR in at least 90% of observations by each reader and superior to FBP in at least 95% of observations by each reader.
In contrast, MBIR was significantly not superior with respect to the perceived image quality of the intrarenal collecting system (Table 2), and the corresponding interobserver agreement was only moderate, with k = 0.75 (95% confidence interval [CI] = 0.48–0.92) for MBIR versus ASIR and k = 0.64 (95% CI = 0.32–0.83) for MBIR versus FBP. Of note, ‘‘significantly not superior’’ is not equivalent to ‘‘significantly inferior’’ and, rather, indicates that MBIR was most frequently scored as equivalent to ASIR and/or FBP in observations of this imaging feature. Overall Perceived Image Quality
Nonlinear least-squares regression analysis was performed to evaluate the relationship between milliamperage and mean overall perceived image quality score as assessed within the kidneys and the adjacent retroperitoneum (Fig 5). The data were well fit by a Gompertz function of the form. Image quality ¼ b1 exp ð exp ½ b2 fmA b3gÞ where b1, b2, and b3 are parameters with estimated values and 95% CIs as displayed in Table 4. Parameters b1 and b3 were 777
OLCOTT ET AL
Academic Radiology, Vol 21, No 6, June 2014
Figure 2. Axial CT images acquired at 255 mA (a–c) and 45 mA (d–f) and reconstructed with MBIR (a, d), ASIR (b, e), and FBP (c, f). Perceived image quality of hilar structures (arrowheads) and focal water-density lesion (arrows) was superior with MBIR at both milliampere levels. Streak artifact is absent in the MBIR images but present in the ASIR and FBP images. Nodular appearance of the lower dose MBIR image (d) compared to the higher dose image (a) is noted. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered backprojection; MBIR, model-based iterative reconstruction.
significantly different for MBIR compared to those for ASIR and FBP (both with P < .001), whereas parameter b2 did not differ significantly (P = .132 and .650, respectively). The MBIR algorithm required less milliamperage to reach any given overall perceived image quality level compared to ASIR and FBP and, similarly, achieved a higher level of overall perceived image quality across the range of tube current levels (Fig 5). The MBIR curve intersected the overall perceived image quality levels of 2 (diagnostic) and 1 (minimally diagnostic) at approximately 105 and 25 mA (by extrapolation), respectively, whereas the ASIR curve did the same at approximately 235 and 100 mA, respectively. The FBP curve behaved similarly to the ASIR curve but followed a somewhat lower course. Thus, MBIR required approximately 130 fewer milliamperes or 55.3% (1 105/235) less radiation dose than ASIR to produce an overall perceived image quality score of 2 (diagnostic) and approximately 75 fewer milliamperes or 75.0% (1 25/100) less dose than ASIR and FBP to achieve an overall perceived image quality score of 1 (minimally diagnostic). Moreover, at all milliampere levels the MBIR curve consistently exceeded the ASIR and FBP curves by 0.8–0.9 points in overall perceived image quality, and only 778
the MBIR curve approached or met the overall perceived image quality score of 3 (superior) at any milliampere level. With respect to overall perceived image quality scores, pairwise agreement among the four readers was moderate for all pairings except those involving reader 3 whose associations were less strong, as shown by the linearly weighted kappa statistic (Table 5). Table 6 indicates the minimum tube current levels at which all four individual readers reported overall perceived image quality scores of at least 1, at least 2, and at least 3 for ASIR, FBP, and MBIR. Of particular note, MBIR and ASIR required 135 and 270 mA, respectively, for scores of at least 2 (diagnostic). The values for FBP were equivalent to those of ASIR. All reviewers described the perceived image quality in lower-milliamperage MBIR images as ‘‘blotchy’’ or ‘‘nodular’’ compared to higher-milliamperage MBIR images whereas lower-milliamperage ASIR and FBP images were described as having familiar ‘‘point-like’’ patterns of noise compared to corresponding higher-milliamperage images (Figs 1,2). Image Noise
Nonlinear least-squares regression analysis was performed to evaluate the relationship between milliampere and mean
Academic Radiology, Vol 21, No 6, June 2014
MBIR OF KIDNEYS AND RETROPERITONEUM
Figure 3. Axial CT images acquired at 225 mA (a–c) and 75 mA (d–f), reconstructed with MBIR (a, d), ASIR (b, e), and FBP (c, f). Perceived image quality of a small urinary calculus (arrows) was equivalent among MBIR, ASIR, and FBP at the higher milliampere level and superior for MBIR at the lower milliampere level. Streak artifact is absent in the MBIR images but present in the ASIR and FBP images. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered back-projection; MBIR, model-based iterative reconstruction.
image noise as assessed within the aorta (Fig 6, Table 7). These data fit a power function of the form Noise ¼ b0 ðmAÞ ^b1 where b0 and b1 are parameters whose estimated values and 95% CIs are shown in Table 7. Both parameters b0 and b1 differed significantly between MBIR and ASIR and between MBIR and FBP (b0: P = .011 and .006, respectively; b1: P < .001 in both cases). Image noise for ASIR and FBP rose with decreasing milliamperes at a rate proportional to nearly the traditionally expected rate of (mA)0.5 (9,12) as evidenced by their b1 values of 0.53 and 0.51, respectively. In contrast, noise for MBIR rose with decreasing milliamperes at a much slower rate, with b1 = 0.22.
DISCUSSION This prospective investigation supports earlier reports that modern iterative image reconstruction algorithms such as MBIR should provide improved CT image quality when compared to images reconstructed with earlier algorithms
such as ASIR or FBP at given radiation dose levels (7–15). Similarly, our results indicate that algorithms such as MBIR produce a given level of perceived image quality with lower radiation dose compared to algorithms such as ASIR and FBP. Our conclusions are based on examinations of the kidneys and the adjacent retroperitoneum as visualized in the renal excretory phase of contrast enhancement. Our study design included comparisons of MBIR to ASIR and FBP along a stepwise gradation of dose levels, from tube current values of 30–300 mA (corresponding CTDIvol 1.4–14 mGy), with evaluation of overall perceived image quality in a stepwise fashion from a score of ‘‘nondiagnostic’’ through a score of ‘‘superior.’’ This design permits a more informative evaluation than one that just compares the clinical adequacy of image quality of one technology operating at a ‘‘typical’’ clinical dose to that of a newer technology operating at a lower dose, with no gradations of dose in between, such that apparent ‘‘dose saving’’ may result simply from a ‘‘typical’’ clinical dose that is higher than necessary to provide adequate image quality. Thus, our data permit a finely scaled comparison of the perceived image quality of MBIR compared to that of ASIR and FBP, as a function of dose. As shown by nonlinear least-squares regression, the MBIR algorithm produced 779
OLCOTT ET AL
Academic Radiology, Vol 21, No 6, June 2014
Figure 4. Axial CT images acquired at 270 mA (a–c) and 30 mA (d–f), reconstructed with MBIR (a, d), ASIR (b, e), and FBP (c, f) and displayed on wide window settings. Perceived image quality of the contrast medium–opacified intrarenal collecting system (arrows) was equivalent among MBIR, ASIR, and FBP at both milliampere levels. Streak artifact is absent in the MBIR images but present in the ASIR and FBP images. ASIR, adaptive statistical iterative reconstruction; CT, computed tomography; FBP, filtered back-projection; MBIR, model-based iterative reconstruction.
Figure 5. Graph showing nonlinear least-squares regression relating tube current (mA) and mean perceived image quality scores for reconstructions performed with MBIR, ASIR, and FBP. Curves describe a Gompertz function of the form ‘‘Image quality = b1 exp (exp [b2 {dose b3}])’’. Vertical bars correspond to + and one standard error from the mean image quality score at each milliampere value. ASIR, adaptive statistical iterative reconstruction; FBP, filtered back-projection; MBIR, model-based iterative reconstruction.
780
significantly better overall perceived image quality than ASIR and FBP did at all dose levels (Fig 5, Table 4). To produce scans of ‘‘diagnostic’’ overall perceived image quality, MBIR required approximately 105 mA or approximately 130 fewer mA than ASIR required, which corresponds to 55.3% (1 105/235) less dose than ASIR required. Similarly, although an observation that is arguably less rigorous in nature, the minimum milliampere levels at which all four individual readers reported overall perceived image quality scores of at least ‘‘diagnostic’’ were 135 mA for MBIR and 270 mA for both ASIR and FBP (Table 6). These findings suggest that a ‘‘floor’’ or minimum dose necessary to produce scans with diagnostic quality in our study exists near 105–135 mA for MBIR, compared to 235–270 mA for ASIR and FBP in our study, representing a dose saving of 50% for MBIR compared to ASIR and FBP. This feature of MBIR may be useful in mitigating medical radiation exposure, a topic of considerable importance (20–26). In addition, MBIR produced superior depiction of several individual imaging features compared to ASIR and FBP and did so significantly when tested against the null hypothesis that MBIR would be scored as superior in 50% of observations
Academic Radiology, Vol 21, No 6, June 2014
MBIR OF KIDNEYS AND RETROPERITONEUM
TABLE 2. Observations of Individual Imaging Features in Which MBIR was Rated Superior to Both ASIR and FBP
Imaging Feature
Percentage of Observations in Which MBIR was Rated Superior
Very Low Dose 30–105 mA 1.4–4.9 mGy
Low Dose 135–225 mA 6.3–10.5 mGy
Standard Dose 255–300 mA 11.9–14.0 mGy
99.2% (123/124) 94.4% (117/124) 96.8% (120/124) 92.5% (74/80)
100.0% (44/44) 97.7% (43/44) 100.0% (44/44) 100.0% (32/32)
97.7% (43/44) 93.2% (41/44) 95.5% (42/44) 91.7% (22/24)
100.0% (124/124) 23.4% (29/124)
100.0% (44/44) 56.8% (25/44) P = .451 100.0% (4/4) P = .125
100.0% (44/44) 6.8% (3/44)
100.0% (36/36) 91.7% (33/36) 94.4% (34/36) 83.3% (20/24) P = .002 100.0% (36/36) 2.8% (1/36)
Retroperitoneal fascia Corticomedullary differentiation Hilar structures Lesions Mitigation of streak artifact Intrarenal collecting system
Percentage of Observations in Which MBIR was Rated Superior, Stratified by Tube Current (mA) and CTDIvol (mGy)
Calculi
6.3% (5/8) P = .727
25.0% (1/4) P = .625
—
ASIR, adaptive statistical iterative reconstruction; FBP, filtered back-projection; MBIR, model-based iterative reconstruction; CTDIvol, volume computed tomography dose index. Numerators indicate observations in which MBIR was rated superior to both ASIR and FBP; denominators indicate the total number of observations. Significance is assessed with the exact binomial test. All P values
