Abstract
BACKGROUND AND PURPOSE: Improved image quality is clinically desired for contrast-enhanced CT of the neck. We compared 30% adaptive statistical iterative reconstruction and model-based iterative reconstruction algorithms for the assessment of image quality of contrast-enhanced CT of the neck.
MATERIALS AND METHODS: Neck contrast-enhanced CT data from 64 consecutive patients were reconstructed retrospectively by using 30% adaptive statistical iterative reconstruction and model-based iterative reconstruction. Objective image quality was assessed by comparing SNR, contrast-to-noise ratio, and background noise at levels 1 (mandible) and 2 (superior mediastinum). Two independent blinded readers subjectively graded the image quality on a scale of 1–5, (grade 5 = excellent image quality without artifacts and grade 1 = nondiagnostic image quality with significant artifacts). The percentage of agreement and disagreement between the 2 readers was assessed.
RESULTS: Compared with 30% adaptive statistical iterative reconstruction, model-based iterative reconstruction significantly improved the SNR and contrast-to-noise ratio at levels 1 and 2. Model-based iterative reconstruction also decreased background noise at level 1 (P = .016), though there was no difference at level 2 (P = .61). Model-based iterative reconstruction was scored higher than 30% adaptive statistical iterative reconstruction by both reviewers at the nasopharynx (P < .001) and oropharynx (P < .001) and for overall image quality (P < .001) and was scored lower at the vocal cords (P < .001) and sternoclavicular junction (P < .001), due to artifacts related to thyroid shielding that were specific for model-based iterative reconstruction.
CONCLUSIONS: Model-based iterative reconstruction offers improved subjective and objective image quality as evidenced by a higher SNR and contrast-to-noise ratio and lower background noise within the same dataset for contrast-enhanced neck CT. Model-based iterative reconstruction has the potential to reduce the radiation dose while maintaining the image quality, with a minor downside being prominent artifacts related to thyroid shield use on model-based iterative reconstruction.
ABBREVIATIONS:
- ASiR30
- 30% adaptive statistical iterative reconstruction
- BN
- background noise
- CNR
- contrast-to-noise ratio
- FBP
- filtered back-projection
- HU
- Hounsfield units
- MBIR
- model-based iterative reconstruction
- PM
- pectoris muscle
- SCM
- sternocleidomastoid muscle
- SVC
- superior vena cava
Since the introduction of CT for medical imaging in the early 1970s, there has been tremendous advancement in overall image quality with concomitant shortening of requisite scan times. Additional major effort has been undertaken to reduce the radiation dose to improve patient safety while maintaining image quality. In particular, image reconstruction algorithms have evolved from the traditional analytic algorithms such as filtered back-projection (FBP) to newer iterative reconstruction methods such as adaptive statistical iterative reconstruction (ASiR; GE Healthcare, Milwaukee, Wisconsin) and most recently model-based iterative reconstruction (MBIR; GE Healthcare), which models system noise statistics and optics.
Both phantom and clinical studies have confirmed that the application of the MBIR algorithm results in an improved contrast-to-noise ratio (CNR), lower background noise (BN),1⇓⇓–4 and reduction of helical conebeam artifacts.2,4 Clinical studies in the delineation of arteries in the posterior fossa on 3D brain CT angiography,1 improved liver lesion detection,3,5 general evaluation of abdominopelvic CT,2 and pediatric chest CT6 all support the use of MBIR, with or without radiation-dose reduction. In this study, we compared objective and subjective image quality in neck CT images reconstructed with 2 different iterative reconstruction algorithms (MBIR versus 30% adaptive statistical iterative reconstruction [ASiR30]) by using the same raw dataset.
Materials and Methods
Patients
The local institutional review board approved this Health Insurance Portability and Accountability Act–compliant retrospective study. Sixty-four consecutive contrast-enhanced CT neck examinations, performed during June and July 2013, were included in this study. There were 42 male (66%) and 22 female (34%) patients (median age, 61 years; age range, 20–84 years). The indications for neck CT examinations included initial head and neck cancer staging, lymphoma staging and surveillance, routine follow-up in the setting of treated tumor, and known or suspected soft-tissue neck infection.
Neck CT Protocol
Details of the contrast-enhanced CT neck protocols are described in Table 1. All CT examinations were performed on a 64–detector row multidetector CT (Discovery HD 75; GE Healthcare), with tube current modulation applied at a noise index of 40 Hounsfield units (HU) for 0.625-mm section thickness by using 0.5-second helical gantry rotation. Because the noise index varies with the section thickness, the predicted noise index for 2.5-mm section thickness was about 23.4 HU.7 For each diagnostic scan, 80 mL of iohexol, 350 mg/mL (Omnipaque 350; GE Healthcare), was intravenously administered via a power injector (the initial 50 mL at a rate of 2 mL/s and the remaining 30 mL at 0.5 mL/s) followed by 30 mL of saline (0.9% NaCl) flush at a rate of 1.5 mL/s. The scan was acquired with a delay of 120 seconds following the commencement of contrast injection. A bismuth thyroid shield was placed over the lower neck in all patients as part of a radiation safety exercise at our institution.8
Soft-tissue neck CT protocol
CT Data Reconstruction
The raw data from each subject's CT examination were reconstructed at the scanner, with a similar magnification, 2.5-mm section thickness at an interval of 2.5 mm in the axial plane, first with ASiR30 and then with MBIR. Details regarding specifically used reconstruction methods were omitted from the generated image sets to ensure complete blinding of the subjective imaging review. The images were sent to the PACS for analysis.
Quantitative Image Analysis
Quantitative image parameters were assessed by a neuroradiologist with a cumulative experience of 8 years (including 2 years of fellowship training in neuroradiology) in interpreting neck CT examinations (S.G.). Several region-of-interest measurements were performed on a PACS workstation by using a 0.5- to 10-mm-diameter circle tool. Mean attenuation values (V) and SD were measured and recorded in Hounsfield units. BN was also measured and recorded as the SD of air. The measurements on each image set were performed at 2 anatomic levels: level 1, the angle of the mandible (Fig 1A) and level 2, the superior mediastinum (Fig 1B). BN was assessed bilaterally at each level, and the mean was calculated. We placed ROIs on the following structures and recorded their respective measurements: The sternocleidomastoid muscle (SCM) (level 1), pectoralis major muscle (PM) (level 2), dominant internal jugular vein (level 1), superior vena cava (SVC) (level 2), and the surrounding fat (both levels) were measured. Signal-to-noise ratio and the CNR were calculated by using the following standard equations9:
Assessment of objective image quality at the level of the mandible (A) and superior mediastinum (B). Region of interest drawn to measure the SD of air bilaterally (background noise), attenuation of the sternocleidomastoid and pectoralis major muscles, internal jugular vein and superior vena cava, and surrounding fat for the estimation of signal-to-noise ratio and contrast-to-noise ratio.
Qualitative Image Analysis
Qualitative image analysis was performed at 4 predefined anatomic levels: 1) the nasopharynx (at the level of fossa of Rosenmuller); 2) the oropharynx (at the level of palatine tonsils); 3) the true vocal cords, and 4) the sternoclavicular junction. In patients with prior tonsillectomy or laryngectomy, images were selected at the level of the surgical bed. On a per-subject basis, identical anatomic levels were selected carefully from both the ASiR30 and MBIR image sets. These 4 anatomic levels were specifically chosen for their increased respective incidence of head and neck cancers and/or a proclivity for streak (due to dental amalgam) or beam-hardening (due to shoulders) artifacts. In addition, readers subjectively evaluated all the images within an image stack (ASiR30 versus MBIR) to assess overall image quality.
Two experienced fellowship-trained neuroradiologists with 9 (J.B. [reader 1]) and 10 (J.B.A. [reader 2]) years of cumulative experience in interpreting neck CT examinations were blinded to the patient information and reconstruction methodology. On a per-patient basis, image stacks obtained from the 2 reconstruction algorithms were displayed side-by-side on each of 2 PACS monitors. Images were presented to reviewers with a prespecified window width and level of 350 and 50 HU, respectively, though reviewers were free to vary both at their discretion. In the absence of a focal lesion, the readers (1 and 2) were asked to evaluate the appearance of normal structures, with special attention paid to the delineation of fat planes, the internal architecture of soft tissues, and overall artifacts burden. In the presence of a focal lesion, readers were asked to evaluate lesion conspicuity, the margins, internal architecture of the lesion, and overall artifacts burden.
Overall image quality and image quality at each of the 4 preselected anatomic locations were graded on a scale of 1–5 (5 = excellent, no artifacts; 4 = good, minimal artifacts; 3 = acceptable, mild artifacts; 2 = suboptimal, moderate artifacts; and 1 = nondiagnostic, significant artifacts). Grades 1 and 2 were considered unacceptable for clinical interpretation. On the basis of the grades of qualitative assessment by each of the readers (1 and 2), the data were divided into 2 categories: category I (readers agree that ASiR>MBIR) and category II (readers agree that MBIR≥ASiR).
Statistical Analysis
The paired Student t test was used to assess differences in objective measures of image quality between the 2 reconstruction algorithms. Histograms were visually inspected to ensure that distributional assumptions were reasonable. The Wilcoxon signed rank test was used to test for differences in subjective image-quality ratings between the 2 reconstruction algorithms and readers. The percentage agreement of the 2 readers (1 and 2) for the categories I (readers agree that ASiR>MBIR) or II (readers agree that MBIR≥ASiR) was assessed at all 4 levels and for overall image quality. All statistical calculations were conducted with the statistical computing language R (R statistical computing software; (http://www.r-project.org/). Throughout, 2-tailed tests were used with P < .05 denoting statistical significance.
Results
Quantitative Analysis
Objective image-quality measurement data are summarized in Table 2. At level 1, the mean BN was significantly lower (P = .016) for MBIR compared with ASiR30. The mean attenuation values of the internal jugular vein and SCM for the MBIR dataset were significantly lower (P = .03 and P < .001, respectively) in comparison with the ASiR30 dataset. At level 2, there was no statistically significant difference (P = .61) in the BN between the MBIR and ASiR30 image sets, whereas the mean attenuation values of SVC and PM for the MBIR dataset were significantly higher (P = .03 and P < .021, respectively) compared with the ASiR30 dataset.
Quantitative assessment of image quality (N = 64 subjects)
The measured SNRs of the SCM and PM were significantly higher (P < .001 and P < .001, respectively) for the MBIR dataset compared with the ASiR30 dataset, likely due to reduced BN. The measured CNR of the internal jugular vein and SVC in relation to the SCM and PM was also significantly higher (P < .001 and P < .001, respectively) for MBIR compared with ASiR30.
Qualitative Analysis
Subjective image-quality assessment is summarized in Table 3. Both readers preferentially graded MBIR over ASiR30 for overall image quality (the entire stack of images) (P < .001) and at the levels of the nasopharynx and oropharynx (Figs 2 and 3). Readers preferentially graded ASiR30 over MBIR (P < .001) at the level of the vocal cords and sternoclavicular region (Fig 4).
Qualitative assessment of image quality, based on average ratings of 2 readers (N = 64 subjects)
Axial images of contrast-enhanced neck CT at the level of oropharynx reconstructed by using 30% adaptive statistical iterative reconstruction (A) and model-based iterative reconstruction (B). Note better conspicuity and definition of tumor margins on MBIR (arrows) compared with ASiR30 (arrows).
Axial images of contrast-enhanced neck CT at the level of oropharynx reconstructed by using 30% adaptive statistical iterative reconstruction (A) and model-based iterative reconstruction (B). Note a decrease in the streak artifacts related to dental amalgam particularly in the region of the base of the tongue, parotids, and foramen magnum and a poorer definition of fat planes in the parapharyngeal space on MBIR images.
Axial images of contrast-enhanced neck CT at the level of sternoclavicular junction reconstructed by using 30% adaptive statistical iterative reconstruction (A) and model-based iterative reconstruction (B). Artifacts are due to a thyroid shield in the display FOV specific for MBIR (arrows). Note the absence of these thyroid shield artifacts on images reconstructed by using ASiR30 (arrows).
The percentage of agreement and disagreement between the readers for categories I and II is presented as a bar graph (Fig 5). Reader agreement in preferring MBIR to ASiR30 was 83% for overall image quality, 100% at the level of nasopharynx, and 98% at the level of the oropharynx. The percentage of agreement between the 2 readers at the level of the vocal cords and sternoclavicular junction, however, was only 41% and 42%, respectively.
Bar graph showing the percentage of agreement and disagreement between the 2 independent readers (1 and 2) for the assessment of subjective image quality at various levels and overall image quality in both categories (category I: ASiR>MBIR and category II: MBIR≥ASiR).
Discussion
CT, enabled by technologic advances and widespread availability, has become the imaging technique of choice in the assessment of numerous neck pathologies. Multidetector CT technology has brought a commensurate increase in the number of CT studies performed,10 leading to a significant increase in the radiation dose related to CT scanning.11,12 Although CT scanning is only approximately 15% of all radiologic examinations, it accounts for up to 70% of the radiation dose to the patient.13 Increasing media attention and public awareness regarding the potential risks of radiation exposure, particularly radiation-induced carcinogenesis, have encouraged development of strategies to further reduce the patient radiation dose. The radiation risk to children is of particular concern because the estimated lifetime cancer risk for a 1-year-old child from the radiation exposure of a head CT is 0.07%.14 Radiation-induced carcinogenesis may be related to a linear no-threshold stochastic effect15 or a cumulative radiation effect,16 particularly in patients with a known history of cancer or other chronic head and neck diseases warranting repeat CT imaging.
There are several strategies for radiation-dose reduction in clinical CT examinations, including changing the CT acquisition parameters (tube current, tube rotation time, peak voltage, pitch, and collimation), use of tube current modulation,17 automatic exposure control,18 adjusting the kV on the basis of patient size,19 iterative reconstruction,20⇓–22 and selective in-plane shielding23 (thyroid, eye lens, breast, or gonadal shield). Developing appropriate CT protocols with an optimal compromise between diagnostic image quality and radiation dose is a team effort of technologist, physicist, and radiologist.24
The FBP algorithm for CT image reconstruction has been the preferred reconstruction algorithm for CT images since its introduction in the 1970s. FBP is based on the assumption that the CT system is perfect and noise-free. However, FBP actually amplifies the quantum and electronic noise in the projection data.25 Its main advantages include speed of image reconstruction due to single-pass direct calculation and production of CT images that are routinely considered clinically acceptable. To overcome the limitation of analytic FBP, iterative techniques for image reconstruction were introduced. ASiR (GE Healthcare) is a partially iterative method using a statistical model of noise that has been commercially available since 2008.20,21 It uses the information obtained by initial FBP reconstruction and then repeatedly compares the estimated pixel value with the ideal value predicted by the noise model, until the estimated and ideal values converge.26
MBIR is a fully iterative method that not only models the statistics of noise (photon statistic and electronic noise) but additionally models the system optics (detector response to incident x-ray beam).4,27 Prior research has suggested that the MBIR algorithm generates CT images with better noise suppression, spatial resolution, conspicuity, and overall image quality compared with FBP, while maintaining similar uniformity and beam-hardening.28 Several studies have reported clinically relevant radiation-dose reduction by using MBIR while still preserving diagnostic image quality.2,5,6,25 One notable drawback of the current version of the MBIR is that it can only reconstruct CT images by using a standard soft-tissue kernel and not a bone kernel; hence, this algorithm fails to provide images with requisite edge details for better assessment of bony lesions.
In this study, the neck CT images reconstructed by using MBIR showed significantly improved SNR and CNR while reducing the BN compared with images reconstructed by using ASiR30. The qualitative assessment by 2 blinded independent readers suggests that MBIR is superior to ASiR30 at the levels of nasopharynx and oropharynx. However, qualitative assessment at the levels of the vocal cords and sternoclavicular region suggests that ASiR30 is preferred over MBIR in these regions. This observation may, in part, be explained by the presence of a bismuth thyroid shield used for radiation-dose reduction to the thyroid gland. The bismuth shield resulted in artifacts projecting over the vocal cords and sternoclavicular region, which was more prominent on images reconstructed with MBIR, particularly when the shield was close to the skin surface. The exact cause of these artifacts in CT images reconstructed with the MBIR algorithm remains unknown, though we believe that it is at least partly related to the current noise model of the MBIR algorithm, which is not well-equipped to deal with the sharp changes in local attenuation and statistics in the projection sinogram produced at the edges of the bismuth shield. We also believe that these thyroid shield artifacts contributed to the underperformance of MBIR at the level of the vocal cords and sternoclavicular region.
It has been reported from various clinical studies that MBIR methods are better than analytic FBP and ASiR algorithms in providing acceptable image quality in the setting of a lower radiation dose.1⇓–3,29⇓–31 The major advantages of MBIR over ASiR include further improvement in CNR, decrease in BN, and decrease in artifacts while preserving the diagnostic image quality.1⇓–3,5,6 In this study, overall image quality in neck CT examinations was improved with MBIR, potentially allowing the exchange of improved image quality for further radiation-dose reduction. A recent study applying MBIR to chest CT reported radiation-dose reductions approaching 70%–80%.16 In contrast, the expected dose reduction from using a bismuth thyroid shield is 28%.8 If a similar radiation-dose reduction can be achieved in contrast-enhanced neck CT examinations by applying a particular postprocessing algorithm, then one could argue that the use of a thyroid shield is less warranted.
The main drawback of MBIR in its present form over other reconstruction algorithms is the computing time, which is about 1 image per second as opposed to 15 images per second for FBP and 10 images per second for ASiR50. In our study, the average reconstruction time for individual soft-tissue neck CT examinations was approximately 45 minutes (range, 30–75 minutes). This time constraint for MBIR, however, should be reduced in the near future due to advancements in computational power.
Study Limitations
Our study has many limitations. First, we did not perform a quality comparison of the MBIR image algorithm with and without the use of a thyroid shield to more accurately compare the image quality of ASiR and MBIR. Second, the qualitative image analysis was a simultaneous assessment of image quality and artifacts burden. As such, this may explain lower subjective scores in areas covered by the thyroid shield. Third, patient data collection was performed in a consecutive fashion, without discrete inclusion criteria to evaluate a specific pathology; thus our study cannot assess the impact of MBIR on the detection of disease.
Conclusions
MBIR offers improved subjective and objective image quality for contrast-enhanced neck CT, suggesting that MBIR may further reduce the radiation dose while maintaining diagnostic image quality. Further studies are necessary to assess how much radiation-dose savings can be achieved with the application of newer and faster model-based iterative reconstruction algorithms, and further development will be required to bring image processing into clinically reasonable timeframes.
Footnotes
Disclosures: Daniel S. Hippe—UNRELATED: GE Healthcare, Philips Healthcare, Comments: research grants outside current work. Also, Society of Interventional Radiology Foundation award for statistical analysis work on another study.
Paper previously presented as an abstract at: American Society of Neuroradiology Annual Meeting and the Foundation of the ASNR Symposium, May 17–22, 2014; Montreal, Quebec, Canada.
REFERENCES
- Received May 25, 2014.
- Accepted after revision August 18, 2014.
- © 2015 by American Journal of Neuroradiology
