ORIGINAL ARTICLE
Comparison of Iterative Model–Based Reconstruction Versus Conventional Filtered Back Projection and Hybrid Iterative Reconstruction Techniques: Lesion Conspicuity and Influence of Body Size in Anthropomorphic Liver Phantoms Jeong Hee Yoon, MD,* Jeong Min Lee, MD,*† Mi Hye Yu, MD,‡ Jee Hyun Baek, MD,§ Ju Hyun Jeon, MD,* Bo Yun Hur, MD,k Amar Dhanantwari, PhD,¶ Se Young Chung, MD,# Joon Koo Han, MD,*† and Byung Ihn Choi, MD*† Purpose: This study aimed to determine whether an iterative model–based reconstruction (IMR) can improve lesion conspicuity and depiction on computed tomography (CT) compared with filtered back projection (FBP) and hybrid iterative reconstruction (iDose4) using anthropomorphic phantoms. Materials and Methods: One small and one large anthropomorphic body phantoms, each containing 8 simulated focal liver lesions (FLLs), were scanned using a 256-channel CT scanner at 120 kVp with variable tube current-time products (10-200 mAs). Scans were divided into 3 groups based on radiation dose (RD) as follows: (a) full dose (FD), (b) low dose (FD50), and (c) ultralow dose (FD25 for the large phantom, FD15 for the small phantom). All images were reconstructed using FBP, iDose4, and IMR. Image noise and lesion-to-liver contrast were assessed quantitatively and qualitatively. Thereafter, 6 radiologists independently evaluated conspicuity of FLLs, and then, compared the number of invisible FLLs on 3 image sets of each RD group. Results: Image noise was significantly lower with IMR than with FBP and iDose4 at the same RD. Iterative model–based reconstruction improved conspicuity of low-contrast FLLs in all RD groups compared to the others (P < 0.001). Furthermore, compared to FBP and iDose4, the number of visible FLLs significantly increased on IMR images in the FD15 group of the small phantom 52.8% [38/72], 68.1% [49/72], and 84.8% [61/72], respectively; P < 0.001) and in the FD 25, FD50 groups of the large phantom (FD50: 56.9% [41/72], 76.4% [55/72], and 84.7% [61/72], respectively; P < 0.05). Conclusions: Iterative model–based reconstruction reduced image noise and improved low-contrast FLL conspicuity, compared to FBP and iDose4. Therefore, depiction of low-contrast FLLs on FBP could be improved using IMR. Key Words: computed tomography, iterative reconstruction, model-based iterative reconstruction, radiation dose, low contrast (J Comput Assist Tomogr 2014;38: 859–868)
W
ith rapid advances in the technical development of multidetector computed tomography (MDCT), the application of
From the *Department of Radiology, Seoul National University Hospital; †Institute of Radiation Medicine, Seoul National University College of Medicine; ‡Department of Radiology, Konkuk University Hospital; §Human Medical Imaging and Intervention Center; kDepartment of Radiology, Seoul National University Boramae Hospital, Seoul, Korea; ¶Philips Healthcare, Cleveland, OH; and #Department of Radiology, Healthcare System Gangnam Center, Seoul National University Hospital, Seoul, Korea. Received for publication May 25, 2014; accepted July 21, 2014. Reprints: Jeong Min Lee, MD, Department of Radiology and Institute of Radiation Medicine, Seoul National University College of Medicine, 101 Daehak-ro, Jongno-gu, Seoul 110-744, Korea (e‐mail: [email protected]). Supplemental digital contents are available for this article. Direct URL citations appear in the printed text, and links to the digital files are provided in the HTML text of this article on the journal’s Web site (www.jcat.org). A.D. is an employee of Philips Healthcare. For the remaining authors, no conflicts of interest were declared. Copyright © 2014 by Lippincott Williams & Wilkins
computed tomography (CT) has seen remarkable increases in the evaluation of various abdominal diseases. This has been particularly true in liver imaging, with CT now widely accepted as the primary imaging tool for the detection and characterization of focal liver lesions (FLLs). However, there have also been increasing concerns over radiation exposure along with the more frequent use of CT.1–3 Several strategies have been suggested to decrease the radiation dose (RD) related with the use of diagnostic CTexaminations in response to these concerns.4–6 One such strategy is the iterative reconstruction (IR) technique which lowers the RD of MDCT through the lowering of image noise.7–9 Among the variously developed IR techniques, model-based IR, which is a full IR technique,10 has shown RD reductions of up to 50% without degradation in image quality.11,12 However, the balance between a reduction in RD and diagnostic performance is still under investigation as a decrease in image quality is, at present, inevitable in low-dose CT scans.13,14 This is especially the case in several pathologic conditions of abdominal CT scans, where both RD reduction and maintenance of optimal image quality are considered essential because of their low lesion-to-background contrast ratios15 compared to most pathologies which show high contrast between the lesion and lung parenchyma in chest CT. Furthermore, it is also known that image quality can degrade in patients with a large body habitus.16 Therefore, this has also been an issue of increasing concern as too much RD reduction even after application of IR techniques may hamper the diagnostic performance of CT in detecting abdominal lesions in these patients.15,17 Recently, a new full IR algorithm (iterative model–based reconstruction [IMR]; Philips Healthcare) which incorporates system optics as well as photon and noise statistics has been developed. Although some researchers anticipate that low-dose abdomen CT using IR algorithms may worsen diagnostic performance,13,14 others believe that diagnostic performance for low-contrast lesions can be improved by reducing noise and preserving edge definitions of the lesions.15,18,19 Until now, however, there have been no studies that have determined how far this prototype full IR algorithm can lower RD without compromising diagnostic performance for the evaluation of FLLs compared with standard filtered back projection (FBP) or hybrid IR algorithms. Therefore, the purpose of this study is to determine whether the newly developed full IR algorithm can reduce RD without compromising lesion conspicuity in small and large body sized phantoms in comparison with FBP and hybrid IR algorithms (iDose4).
MATERIALS AND METHODS Phantoms A customized anthropomorphic phantom was used in this study. This phantom was originally based on the anthropomorphic
J Comput Assist Tomogr • Volume 38, Number 6, November/December 2014
www.jcat.org
Copyright © 2014 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
859
J Comput Assist Tomogr • Volume 38, Number 6, November/December 2014
Yoon et al
PH-5 phantom (Kyoto Kagaku, Japan) which mimics the human torso including the liver, pancreas, kidneys, spleen, inferior vena cava, and aorta during the portal venous phase. Attenuation of the organs in this customized phantom was verified by comparing human data in the preliminary study with the phantom during its construction. The original phantom, with a diameter of 25 cm, was used to emulate a small body habitus (Fig. 1A); thereafter, we tightly wrapped layers of pork belly fat with a thickness of 5 cm around the phantom (diameter of 35 cm, Fig. 1B) to mimic a subject with a large body habitus. In the liver of the phantom, eight 15-mm sized FLLs were implemented. The implemented FLLs consisted of 4 hypoattenuating lesions with different degrees of lesion-to-liver contrast (−10, −20, −30, and −50 HU) and 4 hyperattenuating lesions also with variable degrees of lesion-to-liver contrast (+10, +20, +30, and +50 HU). On the basis of the attenuation, FLLs were classified into 2 groups as follows: high-contrast FLLs (±30 and ±50 HU) and low-contrast FLLs (±10 and ±20 HU).
CT Image Acquisition Computed tomographic scans of the anthropomorphic phantom were obtained using a CT scanner equipped with 256-row detectors (iCT256, Philips HealthCare). The phantom was placed at the isocenter of the CT scanner with its cross section perpendicular to the scanner's z axis so as to avoid unnecessary noise.20 In all scans, the following parameters were kept constant: detector collimation, 128 0.625 mm; rotation time, 0.5 seconds; pitch, 1.0; fields of view, 350.0 mm (small) or 500.0 mm (large); and slice thickness, 3 mm. Multiple CT scans were obtained under the standard tube voltage (120 kVp) and 9 different tube currentstime products (20, 40, 60, 80, 100, 130, 150, 180, and 200 mAs) for the 2 phantom models and additional 10 mAs for a small phantom. Automatic tube current modulation was not used to keep RD constant. On the basis of tube current-time products, CT scans were divided into 3 RD groups in the small phantom as follows: (a) full-dose (FD) group: 200, 180, and 150 mAs (177 mAs on average); (b) low-dose (FD50) group: 100, 80, and 60 mAs (80 mAs on average, 54.8% dose reduction); and (c) ultralow-dose ( FD15 group: 40, 20, and 10 mAs (23.3 mAs on average, 86.8% dose reduction). In the large phantom, CT scans were also categorized into 3 groups in the same manner: (a) FD group: 200 and 180 (190 mAs on average); (b) FD50 group: 130, 100, and 80 mAs (103.3 mAs on average, 45.6% dose reduction); and (c) FD25 group: 60, 40, and 20 mAs (40 mAs on average, 78.9% dose reduction). Different dose settings in ultralow RD groups (FD15 and FD25) in both phantoms were determined based on the
currently used automatic dose modulation technique, which may cause different RDs depending on patients’ body size.
CT Image Reconstruction Scanned data were reconstructed with FBP using a soft tissue kernel, a hybrid IR algorithm with a strength of 4 (iDose4), and a prototype IMR algorithm in the level 2 low-contrast mode (IMR L2). iDose4 uses both imaging and projection data to identify noise and proceeds iteration in both domains, and iteration level can be individualized with user-selected levels.21 Iterative model– based reconstruction, which is a full IR method of the same vendor, adding information of system optics, is discussed later in detail.
Iterative Model–Based Reconstruction Iterative model–based reconstruction algorithm consists of data statistics, image statistics, and system optics. In data statistics, the optimization process is performed by maintaining data fidelity and penalizing noise: photon statistics and electronic noise are considered to optimize the projection data.22 In image statistics, noise penalty and image accuracy are performed. In addition, system optics of the CT scanner (iCT256) including the real size of focal spots and detectors is adapted. With the estimated noise statistics, system optics can formulate a mathematical model, and the reconstructed image is then compared with the acquired data until the 2 data approaches the best match. In this algorithm, the penalized maximum log-likelihood algorithm is used to maintain true data. The edge-preserving function is performed to maintain edges between the 2 different attenuations. The algorithm consists of 3 settings taking into consideration the objects of interests: (a) the low contrast setting (Soft tissue) is indicated for low-contrast lesions, usually abdominopelvic CT; (b) the sharp setting (SharpPlus) is for high-contrast regions such as the lung, or sinonasal or temporal bones; and (c) the combo setting (Routine) targets CT angiography. In each setting, the noise reduction level can be adjusted to 3 levels as follows: level 1 (L1), low level of noise reduction; level 2 (L2), medium level of noise reduction; and level 3 (L3), high level of noise reduction. Reconstruction time is approximately 15 minutes to complete 1 abdomen CT reconstruction using dedicated parallel processors on a prototype system. In this study, based on our preliminary study results, which compared the image quality and diagnostic acceptability of IMR L1, L2, and L3, the level 2 IMR technique was selected (Appendix 1).
FIGURE 1. Appearance of the customized phantom. Appearance of the small phantom, made based on a commercially available phantom (PH-5, Kyoto Kagaku) (A). To mimic a human with a large body habitus, the phantom was tightly wrapped with layers of pork belly with a thickness of 5 cm (B). Figure 1 can be viewed online in color at www.jcat.org.
860
www.jcat.org
© 2014 Lippincott Williams & Wilkins
Copyright © 2014 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
J Comput Assist Tomogr • Volume 38, Number 6, November/December 2014
RD Measurement The volume CT dose index (CTDIvol) and the dose length product (DLP) were recorded for each RD group of the 2 phantoms. Effective dose (ED) was calculated by multiplying the tissue conversion factor of the abdomen (0.017).23
Image Analysis Computed tomographic data were reviewed by 5 reviewers in a blinded manner on high-resolution monitors with a spatial resolution of 1600 1200 (Totoku, Japan). The reviewers (J.H.Y. with 8 years of clinical experience; M.H.Y. with 9 years of clinical experience; J.J.H. with 9 years of clinical experience; J.H.B. with 11 years of clinical experience; and B.Y.H. with 6 years of clinical experience) were allowed to change the window level and width during the session, as in daily practice. Computed tomographic scan data were randomly distributed to reviewers and scan information such as tube current-time products, and reconstruction algorithms were blinded.
Quantitative Analysis Computed tomographic attenuation of the liver and the 8 FLLs as well as image noise were measured for the CT image sets reconstructed with the 3 different algorithms (FBP, iDose4, and IMR). To evaluate image noise, 3 circular regions of interests (ROIs) were drawn in the anterior abdominal wall and both paraspinal muscles (mean, 170.9 mm2; range, 120.6–225.4 mm2) by an attending radiologist (J.H.Y.) and the average values were determined as image noise since the phantom originally does not have a subcutaneous fat layer, and the abdominal and paraspinal muscles were made of the same, homogeneous material. To measure attenuation of the parenchyma of the liver in the phantom, 3 circular ROIs were drawn for the liver left lateral segment, right anterior segment, and right posterior segment (mean, 453.3 mm2; range, 300.5–597.0 mm2). To ensure consistency, all measurements were performed 3 times and the average values of the attenuation coefficients were used. Attenuation coefficients of the 8 tumors in the phantoms were measured by drawing ROIs as large as possible so as to encompass the whole lesions as much as possible (mean, 86.9 mm2; range, 40.4–121.1 mm2). To avoid misregistration between the ROI and tumors among the image sets reconstructed
IMR vs Conventional FBP and Hybrid IR Techniques
with the 3 different reconstruction algorithms, ROIs were copied and pasted in all scans in the same RD group. For each of the image sets, the lesion-to-liver contrast-to-noise ratio (CNR) was calculated using the following equation: CNR = │ROIlesion − ROIliver│/N, where ROIlesion is the mean HU of the tumors, ROIliver is the mean HU of the liver parenchyma, and N is image noise.
Qualitative Analysis Before qualitative analysis, 3 attending abdominal radiologists (J.H.Y., M.H.Y., and J.M.L. with 22 years of abdominal CT imaging) established criteria for image noise, lesion conspicuity, and diagnostic acceptability in consensus. Image noise was graded on a 5-point scale based on previous studies as follows14,24: score 1, unacceptably noisy; score 2, noisier than average; score 3, average; score 4, less noisy than average; and score 5, minimal or absent noise. Diagnostic acceptability was evaluated as follows: score 1, diagnostically unacceptable; score 2, suboptimal for diagnosis; score 3, average; score 4, better than average; and score 5, excellent—the most preferred. Lesion conspicuity was assessed on a 4-point scale as follows: score 1, not detectable; score 2, barely delineated; score 3, detected with relatively good contrast, but with a blurry margin and presence of contour (not round shape); and score 4, definitely distinct, round shape tumors with a sharp margin and clear contrast. We did not evaluate lesion detectability in this study because all scans were performed using the same phantom so that reviewers were easily aware of FLLs. Instead, the numbers of invisible lesions (score 1) were recorded for each reconstruction algorithm in all RD groups in the 2 phantom models, based on a previous study, which reported relatively low false-positive and high false-negative ratios for FLL detection.13,25 Image noise and diagnostic acceptability of scans with different reconstruction algorithms were analyzed by 2 attending abdominal radiologists (J.H.Y. and M.H.Y.). Reviews were performed independently, followed by a comprehensive review in consensus. In addition, lesion conspicuity was also assessed independently for all 8 tumors on each reconstructed image by 6 abdominal radiologists (J.M.L., J.H.Y., M.H.Y., J.H.J., J.H.B., and B.Y.H.). A total of 432 FLLs in the small phantom (8 FLLs per image set, 9 RDs, 6 radiologists) per reconstruction method and 384 FLLs in the large phantom (8 FLLs per image set, 8 RDs, 6 radiologists) per reconstruction method were analyzed.
TABLE 1. RD and Image Noise of Different Reconstruction Methods at Different RD Reduction Settings in Small and Large Phantoms Small Phantom FD
FD50
RD 11.7 (1.67) 5.28 (1.36) CTDIvol, Gy DLP, mGy·cm 390.9 (55.9) 177.2 (45.4) ED, mSv 6.65 (0.95) 3.01 (0.77) Image noise FBP 13.6 (0.68) 20.1 (2.96) [47.7] 10.1 (0.53) [−25.7] 13.6 (1.47) [0] iDose4 IMR L2 5.5 (0.49) [−59.5] 5.83 (0.35) [−57.1]
Large Phantom FD15 1.55 (1.01) 51.7 (33.9) 0.88 (0.58) 45.0 (18.7) [231] 20.4 (3.17) [50] 7.18 (1.67) [−47.2]
FD
FD50
12.5 (0.94) 420.5 (31.5) 7.15 (0.54)
6.85 (1.67) 229.7 (55.9) 3.90 (0.95)
FD25 2.63 (1.3) 88.3 (43.5) 1.50 (0.74)
25.5 (2.38) 34.95 (5.27) [37.1] 67.6 (22.8) [165.1] 14.5 (0.70) [−43.1] 16.8 (1.82) [−34.1] 21.8 (4.15) [−14.5] 5.20 (0.71) [−79.6] 7.41 (0.48) [−70.9] 9.16 (1.23) [−64]
Values are mean (SD). Numbers in brackets are % of change in image noise in comparison with the image noise of FBP standard mAs in each phantom. FD indicates full-dose (200, 180, and 150 mAs in the small phantom; 200 and 180 mAs in the large phantom); FD15,15% of full-dose (40, 20, and 10 mAs in the small phantom); FD25, 25% of full-dose (60, 40 and 20 mAs in the large phantom); FD50, 50% of full-dose (100, 80, and 60 mAs in the small phantom and 130, 100, 80 mAs in the large phantom); IMR L2 indicates iterative model–based reconstruction level 2.
© 2014 Lippincott Williams & Wilkins
www.jcat.org
Copyright © 2014 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
861
J Comput Assist Tomogr • Volume 38, Number 6, November/December 2014
Yoon et al
TABLE 2. Qualitative Assessment Scores Based on Consensus Review of Image Quality at Different Reconstruction Methods and Different RD Reduction Settings in Small and Large Phantoms Small Phantom
Image noise FBP iDose4 IMR L2 P* FBP vs iDose4 FBP vs IMR L2 iDose4 vs IMR L2 Diagnostic acceptability FBP iDose4 IMR L2 P* FBP vs iDose4 FBP vs IMR L2 iDose4 vs IMR L2
Large Phantom
FD
FD50
FD15
FD
FD50
FD25
3.0 (0.0) 3.6 (0.5) 5.0 (0.0)
Comparison of Iterative Model–Based Reconstruction Versus Conventional Filtered Back Projection and Hybrid Iterative Reconstruction Techniques: Lesion Conspicuity and Influence of Body Size in Anthropomorphic Liver Phantoms Jeong Hee Yoon, MD,* Jeong Min Lee, MD,*† Mi Hye Yu, MD,‡ Jee Hyun Baek, MD,§ Ju Hyun Jeon, MD,* Bo Yun Hur, MD,k Amar Dhanantwari, PhD,¶ Se Young Chung, MD,# Joon Koo Han, MD,*† and Byung Ihn Choi, MD*† Purpose: This study aimed to determine whether an iterative model–based reconstruction (IMR) can improve lesion conspicuity and depiction on computed tomography (CT) compared with filtered back projection (FBP) and hybrid iterative reconstruction (iDose4) using anthropomorphic phantoms. Materials and Methods: One small and one large anthropomorphic body phantoms, each containing 8 simulated focal liver lesions (FLLs), were scanned using a 256-channel CT scanner at 120 kVp with variable tube current-time products (10-200 mAs). Scans were divided into 3 groups based on radiation dose (RD) as follows: (a) full dose (FD), (b) low dose (FD50), and (c) ultralow dose (FD25 for the large phantom, FD15 for the small phantom). All images were reconstructed using FBP, iDose4, and IMR. Image noise and lesion-to-liver contrast were assessed quantitatively and qualitatively. Thereafter, 6 radiologists independently evaluated conspicuity of FLLs, and then, compared the number of invisible FLLs on 3 image sets of each RD group. Results: Image noise was significantly lower with IMR than with FBP and iDose4 at the same RD. Iterative model–based reconstruction improved conspicuity of low-contrast FLLs in all RD groups compared to the others (P < 0.001). Furthermore, compared to FBP and iDose4, the number of visible FLLs significantly increased on IMR images in the FD15 group of the small phantom 52.8% [38/72], 68.1% [49/72], and 84.8% [61/72], respectively; P < 0.001) and in the FD 25, FD50 groups of the large phantom (FD50: 56.9% [41/72], 76.4% [55/72], and 84.7% [61/72], respectively; P < 0.05). Conclusions: Iterative model–based reconstruction reduced image noise and improved low-contrast FLL conspicuity, compared to FBP and iDose4. Therefore, depiction of low-contrast FLLs on FBP could be improved using IMR. Key Words: computed tomography, iterative reconstruction, model-based iterative reconstruction, radiation dose, low contrast (J Comput Assist Tomogr 2014;38: 859–868)
W
ith rapid advances in the technical development of multidetector computed tomography (MDCT), the application of
From the *Department of Radiology, Seoul National University Hospital; †Institute of Radiation Medicine, Seoul National University College of Medicine; ‡Department of Radiology, Konkuk University Hospital; §Human Medical Imaging and Intervention Center; kDepartment of Radiology, Seoul National University Boramae Hospital, Seoul, Korea; ¶Philips Healthcare, Cleveland, OH; and #Department of Radiology, Healthcare System Gangnam Center, Seoul National University Hospital, Seoul, Korea. Received for publication May 25, 2014; accepted July 21, 2014. Reprints: Jeong Min Lee, MD, Department of Radiology and Institute of Radiation Medicine, Seoul National University College of Medicine, 101 Daehak-ro, Jongno-gu, Seoul 110-744, Korea (e‐mail: [email protected]). Supplemental digital contents are available for this article. Direct URL citations appear in the printed text, and links to the digital files are provided in the HTML text of this article on the journal’s Web site (www.jcat.org). A.D. is an employee of Philips Healthcare. For the remaining authors, no conflicts of interest were declared. Copyright © 2014 by Lippincott Williams & Wilkins
computed tomography (CT) has seen remarkable increases in the evaluation of various abdominal diseases. This has been particularly true in liver imaging, with CT now widely accepted as the primary imaging tool for the detection and characterization of focal liver lesions (FLLs). However, there have also been increasing concerns over radiation exposure along with the more frequent use of CT.1–3 Several strategies have been suggested to decrease the radiation dose (RD) related with the use of diagnostic CTexaminations in response to these concerns.4–6 One such strategy is the iterative reconstruction (IR) technique which lowers the RD of MDCT through the lowering of image noise.7–9 Among the variously developed IR techniques, model-based IR, which is a full IR technique,10 has shown RD reductions of up to 50% without degradation in image quality.11,12 However, the balance between a reduction in RD and diagnostic performance is still under investigation as a decrease in image quality is, at present, inevitable in low-dose CT scans.13,14 This is especially the case in several pathologic conditions of abdominal CT scans, where both RD reduction and maintenance of optimal image quality are considered essential because of their low lesion-to-background contrast ratios15 compared to most pathologies which show high contrast between the lesion and lung parenchyma in chest CT. Furthermore, it is also known that image quality can degrade in patients with a large body habitus.16 Therefore, this has also been an issue of increasing concern as too much RD reduction even after application of IR techniques may hamper the diagnostic performance of CT in detecting abdominal lesions in these patients.15,17 Recently, a new full IR algorithm (iterative model–based reconstruction [IMR]; Philips Healthcare) which incorporates system optics as well as photon and noise statistics has been developed. Although some researchers anticipate that low-dose abdomen CT using IR algorithms may worsen diagnostic performance,13,14 others believe that diagnostic performance for low-contrast lesions can be improved by reducing noise and preserving edge definitions of the lesions.15,18,19 Until now, however, there have been no studies that have determined how far this prototype full IR algorithm can lower RD without compromising diagnostic performance for the evaluation of FLLs compared with standard filtered back projection (FBP) or hybrid IR algorithms. Therefore, the purpose of this study is to determine whether the newly developed full IR algorithm can reduce RD without compromising lesion conspicuity in small and large body sized phantoms in comparison with FBP and hybrid IR algorithms (iDose4).
MATERIALS AND METHODS Phantoms A customized anthropomorphic phantom was used in this study. This phantom was originally based on the anthropomorphic
J Comput Assist Tomogr • Volume 38, Number 6, November/December 2014
www.jcat.org
Copyright © 2014 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
859
J Comput Assist Tomogr • Volume 38, Number 6, November/December 2014
Yoon et al
PH-5 phantom (Kyoto Kagaku, Japan) which mimics the human torso including the liver, pancreas, kidneys, spleen, inferior vena cava, and aorta during the portal venous phase. Attenuation of the organs in this customized phantom was verified by comparing human data in the preliminary study with the phantom during its construction. The original phantom, with a diameter of 25 cm, was used to emulate a small body habitus (Fig. 1A); thereafter, we tightly wrapped layers of pork belly fat with a thickness of 5 cm around the phantom (diameter of 35 cm, Fig. 1B) to mimic a subject with a large body habitus. In the liver of the phantom, eight 15-mm sized FLLs were implemented. The implemented FLLs consisted of 4 hypoattenuating lesions with different degrees of lesion-to-liver contrast (−10, −20, −30, and −50 HU) and 4 hyperattenuating lesions also with variable degrees of lesion-to-liver contrast (+10, +20, +30, and +50 HU). On the basis of the attenuation, FLLs were classified into 2 groups as follows: high-contrast FLLs (±30 and ±50 HU) and low-contrast FLLs (±10 and ±20 HU).
CT Image Acquisition Computed tomographic scans of the anthropomorphic phantom were obtained using a CT scanner equipped with 256-row detectors (iCT256, Philips HealthCare). The phantom was placed at the isocenter of the CT scanner with its cross section perpendicular to the scanner's z axis so as to avoid unnecessary noise.20 In all scans, the following parameters were kept constant: detector collimation, 128 0.625 mm; rotation time, 0.5 seconds; pitch, 1.0; fields of view, 350.0 mm (small) or 500.0 mm (large); and slice thickness, 3 mm. Multiple CT scans were obtained under the standard tube voltage (120 kVp) and 9 different tube currentstime products (20, 40, 60, 80, 100, 130, 150, 180, and 200 mAs) for the 2 phantom models and additional 10 mAs for a small phantom. Automatic tube current modulation was not used to keep RD constant. On the basis of tube current-time products, CT scans were divided into 3 RD groups in the small phantom as follows: (a) full-dose (FD) group: 200, 180, and 150 mAs (177 mAs on average); (b) low-dose (FD50) group: 100, 80, and 60 mAs (80 mAs on average, 54.8% dose reduction); and (c) ultralow-dose ( FD15 group: 40, 20, and 10 mAs (23.3 mAs on average, 86.8% dose reduction). In the large phantom, CT scans were also categorized into 3 groups in the same manner: (a) FD group: 200 and 180 (190 mAs on average); (b) FD50 group: 130, 100, and 80 mAs (103.3 mAs on average, 45.6% dose reduction); and (c) FD25 group: 60, 40, and 20 mAs (40 mAs on average, 78.9% dose reduction). Different dose settings in ultralow RD groups (FD15 and FD25) in both phantoms were determined based on the
currently used automatic dose modulation technique, which may cause different RDs depending on patients’ body size.
CT Image Reconstruction Scanned data were reconstructed with FBP using a soft tissue kernel, a hybrid IR algorithm with a strength of 4 (iDose4), and a prototype IMR algorithm in the level 2 low-contrast mode (IMR L2). iDose4 uses both imaging and projection data to identify noise and proceeds iteration in both domains, and iteration level can be individualized with user-selected levels.21 Iterative model– based reconstruction, which is a full IR method of the same vendor, adding information of system optics, is discussed later in detail.
Iterative Model–Based Reconstruction Iterative model–based reconstruction algorithm consists of data statistics, image statistics, and system optics. In data statistics, the optimization process is performed by maintaining data fidelity and penalizing noise: photon statistics and electronic noise are considered to optimize the projection data.22 In image statistics, noise penalty and image accuracy are performed. In addition, system optics of the CT scanner (iCT256) including the real size of focal spots and detectors is adapted. With the estimated noise statistics, system optics can formulate a mathematical model, and the reconstructed image is then compared with the acquired data until the 2 data approaches the best match. In this algorithm, the penalized maximum log-likelihood algorithm is used to maintain true data. The edge-preserving function is performed to maintain edges between the 2 different attenuations. The algorithm consists of 3 settings taking into consideration the objects of interests: (a) the low contrast setting (Soft tissue) is indicated for low-contrast lesions, usually abdominopelvic CT; (b) the sharp setting (SharpPlus) is for high-contrast regions such as the lung, or sinonasal or temporal bones; and (c) the combo setting (Routine) targets CT angiography. In each setting, the noise reduction level can be adjusted to 3 levels as follows: level 1 (L1), low level of noise reduction; level 2 (L2), medium level of noise reduction; and level 3 (L3), high level of noise reduction. Reconstruction time is approximately 15 minutes to complete 1 abdomen CT reconstruction using dedicated parallel processors on a prototype system. In this study, based on our preliminary study results, which compared the image quality and diagnostic acceptability of IMR L1, L2, and L3, the level 2 IMR technique was selected (Appendix 1).
FIGURE 1. Appearance of the customized phantom. Appearance of the small phantom, made based on a commercially available phantom (PH-5, Kyoto Kagaku) (A). To mimic a human with a large body habitus, the phantom was tightly wrapped with layers of pork belly with a thickness of 5 cm (B). Figure 1 can be viewed online in color at www.jcat.org.
860
www.jcat.org
© 2014 Lippincott Williams & Wilkins
Copyright © 2014 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
J Comput Assist Tomogr • Volume 38, Number 6, November/December 2014
RD Measurement The volume CT dose index (CTDIvol) and the dose length product (DLP) were recorded for each RD group of the 2 phantoms. Effective dose (ED) was calculated by multiplying the tissue conversion factor of the abdomen (0.017).23
Image Analysis Computed tomographic data were reviewed by 5 reviewers in a blinded manner on high-resolution monitors with a spatial resolution of 1600 1200 (Totoku, Japan). The reviewers (J.H.Y. with 8 years of clinical experience; M.H.Y. with 9 years of clinical experience; J.J.H. with 9 years of clinical experience; J.H.B. with 11 years of clinical experience; and B.Y.H. with 6 years of clinical experience) were allowed to change the window level and width during the session, as in daily practice. Computed tomographic scan data were randomly distributed to reviewers and scan information such as tube current-time products, and reconstruction algorithms were blinded.
Quantitative Analysis Computed tomographic attenuation of the liver and the 8 FLLs as well as image noise were measured for the CT image sets reconstructed with the 3 different algorithms (FBP, iDose4, and IMR). To evaluate image noise, 3 circular regions of interests (ROIs) were drawn in the anterior abdominal wall and both paraspinal muscles (mean, 170.9 mm2; range, 120.6–225.4 mm2) by an attending radiologist (J.H.Y.) and the average values were determined as image noise since the phantom originally does not have a subcutaneous fat layer, and the abdominal and paraspinal muscles were made of the same, homogeneous material. To measure attenuation of the parenchyma of the liver in the phantom, 3 circular ROIs were drawn for the liver left lateral segment, right anterior segment, and right posterior segment (mean, 453.3 mm2; range, 300.5–597.0 mm2). To ensure consistency, all measurements were performed 3 times and the average values of the attenuation coefficients were used. Attenuation coefficients of the 8 tumors in the phantoms were measured by drawing ROIs as large as possible so as to encompass the whole lesions as much as possible (mean, 86.9 mm2; range, 40.4–121.1 mm2). To avoid misregistration between the ROI and tumors among the image sets reconstructed
IMR vs Conventional FBP and Hybrid IR Techniques
with the 3 different reconstruction algorithms, ROIs were copied and pasted in all scans in the same RD group. For each of the image sets, the lesion-to-liver contrast-to-noise ratio (CNR) was calculated using the following equation: CNR = │ROIlesion − ROIliver│/N, where ROIlesion is the mean HU of the tumors, ROIliver is the mean HU of the liver parenchyma, and N is image noise.
Qualitative Analysis Before qualitative analysis, 3 attending abdominal radiologists (J.H.Y., M.H.Y., and J.M.L. with 22 years of abdominal CT imaging) established criteria for image noise, lesion conspicuity, and diagnostic acceptability in consensus. Image noise was graded on a 5-point scale based on previous studies as follows14,24: score 1, unacceptably noisy; score 2, noisier than average; score 3, average; score 4, less noisy than average; and score 5, minimal or absent noise. Diagnostic acceptability was evaluated as follows: score 1, diagnostically unacceptable; score 2, suboptimal for diagnosis; score 3, average; score 4, better than average; and score 5, excellent—the most preferred. Lesion conspicuity was assessed on a 4-point scale as follows: score 1, not detectable; score 2, barely delineated; score 3, detected with relatively good contrast, but with a blurry margin and presence of contour (not round shape); and score 4, definitely distinct, round shape tumors with a sharp margin and clear contrast. We did not evaluate lesion detectability in this study because all scans were performed using the same phantom so that reviewers were easily aware of FLLs. Instead, the numbers of invisible lesions (score 1) were recorded for each reconstruction algorithm in all RD groups in the 2 phantom models, based on a previous study, which reported relatively low false-positive and high false-negative ratios for FLL detection.13,25 Image noise and diagnostic acceptability of scans with different reconstruction algorithms were analyzed by 2 attending abdominal radiologists (J.H.Y. and M.H.Y.). Reviews were performed independently, followed by a comprehensive review in consensus. In addition, lesion conspicuity was also assessed independently for all 8 tumors on each reconstructed image by 6 abdominal radiologists (J.M.L., J.H.Y., M.H.Y., J.H.J., J.H.B., and B.Y.H.). A total of 432 FLLs in the small phantom (8 FLLs per image set, 9 RDs, 6 radiologists) per reconstruction method and 384 FLLs in the large phantom (8 FLLs per image set, 8 RDs, 6 radiologists) per reconstruction method were analyzed.
TABLE 1. RD and Image Noise of Different Reconstruction Methods at Different RD Reduction Settings in Small and Large Phantoms Small Phantom FD
FD50
RD 11.7 (1.67) 5.28 (1.36) CTDIvol, Gy DLP, mGy·cm 390.9 (55.9) 177.2 (45.4) ED, mSv 6.65 (0.95) 3.01 (0.77) Image noise FBP 13.6 (0.68) 20.1 (2.96) [47.7] 10.1 (0.53) [−25.7] 13.6 (1.47) [0] iDose4 IMR L2 5.5 (0.49) [−59.5] 5.83 (0.35) [−57.1]
Large Phantom FD15 1.55 (1.01) 51.7 (33.9) 0.88 (0.58) 45.0 (18.7) [231] 20.4 (3.17) [50] 7.18 (1.67) [−47.2]
FD
FD50
12.5 (0.94) 420.5 (31.5) 7.15 (0.54)
6.85 (1.67) 229.7 (55.9) 3.90 (0.95)
FD25 2.63 (1.3) 88.3 (43.5) 1.50 (0.74)
25.5 (2.38) 34.95 (5.27) [37.1] 67.6 (22.8) [165.1] 14.5 (0.70) [−43.1] 16.8 (1.82) [−34.1] 21.8 (4.15) [−14.5] 5.20 (0.71) [−79.6] 7.41 (0.48) [−70.9] 9.16 (1.23) [−64]
Values are mean (SD). Numbers in brackets are % of change in image noise in comparison with the image noise of FBP standard mAs in each phantom. FD indicates full-dose (200, 180, and 150 mAs in the small phantom; 200 and 180 mAs in the large phantom); FD15,15% of full-dose (40, 20, and 10 mAs in the small phantom); FD25, 25% of full-dose (60, 40 and 20 mAs in the large phantom); FD50, 50% of full-dose (100, 80, and 60 mAs in the small phantom and 130, 100, 80 mAs in the large phantom); IMR L2 indicates iterative model–based reconstruction level 2.
© 2014 Lippincott Williams & Wilkins
www.jcat.org
Copyright © 2014 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
861
J Comput Assist Tomogr • Volume 38, Number 6, November/December 2014
Yoon et al
TABLE 2. Qualitative Assessment Scores Based on Consensus Review of Image Quality at Different Reconstruction Methods and Different RD Reduction Settings in Small and Large Phantoms Small Phantom
Image noise FBP iDose4 IMR L2 P* FBP vs iDose4 FBP vs IMR L2 iDose4 vs IMR L2 Diagnostic acceptability FBP iDose4 IMR L2 P* FBP vs iDose4 FBP vs IMR L2 iDose4 vs IMR L2
Large Phantom
FD
FD50
FD15
FD
FD50
FD25
3.0 (0.0) 3.6 (0.5) 5.0 (0.0)
