Choosing the Best Reconstruction Technique in · PDF fileChoosing the Best Reconstruction...

ORIGINAL ARTICLE

Choosing the Best Reconstruction Technique in AbdominalComputed Tomography: A Systematic Approach

Hilde Kjernlie Andersen, MSc,* Kristin Jensen, MSc,*† Audun Elnaes Berstad, MD, PhD,‡Trond Mogens Aaløkken, MD, PhD,‡ Joanna Kristiansen, MSc,‡ Bjørn von Gohren Edwin, MD, PhD,*§

Gaute Hagen, MD, PhD,‡ and Anne Catrine Trægde Martinsen, PhD*†

Objective: There is uncertainty regarding the effect of iterative recon-struction (IR) techniques and other reconstruction algorithms on imagequality. The aim of this study was to optimize image quality in relation toradiation dose in computed tomography (CT) liver examinations by com-paring images reconstructed with different abdominal filters with andwithout IR.Methods: An anthropomorphic phantom was scanned on a ToshibaAquilion ONE CT scanner. Images at 2 different dose levels were recon-structed with 12 different body reconstruction filters, all with both filteredback-projection and Adaptive Iterative Dose Reduction 3 dimensional.Receiver operating characteristic curves were constructed. The 2 recon-struction combinations with the highest scores from the phantom studywere evaluated in a second comparison of clinical images. Six liver exam-inations were reconstructed with both filters and evaluated using visualgrading analysis.Results: Two combinations of reconstruction filters and IR were the only2 options among the 8 best images at both dose levels (area under thecurve, 0.96 and 0.94 for 15 mGy as well as 0.86 and 0.84 for 10 mGy).In the patient study, one of these filters in combination with IR scoredslightly higher than the other in combination with IR (mean score, 2.60and 2.57, respectively; P = 0.56). Iterative reconstruction did not signifi-cantly increase lesion detectability for any of the filters.Conclusions: This study indicates that the preferred choice for recon-struction of CT liver examinations performed with the Toshiba AquilionONE should be the FC18 filter with IR, although the IR technique didnot significantly improve lesion detectability and did not compensate forthe dose reduction in this study.

Key Words: CT, Toshiba, iterative reconstruction, reconstruction, ROC,image quality, liver, lesion detectability, optimization

(J Comput Assist Tomogr 2014;00: 00–00)

Computed tomography (CT) is the radiological imaging mo-dality that contributes the most to patient radiation dose from

medical radiation. The United Nations Scientific Committee onthe Effects of Atomic Radiation concluded that there is an in-crease in population dose from medical exposure worldwide,which is mainly a result of the increased annual frequency ofCT scanning.1

Image quality in CT generally improves with increasing radi-ation dose. Recently, there has been a substantial effort from allvendors to improve reconstruction algorithms and image post-processing tools to enhance image quality without generating a

From *The Intervention Centre, Oslo University Hospital; †Department ofPhysics, University in Oslo; ‡Department of Radiology and Nuclear Medicine,Oslo University Hospital, Rikshospitalet; and §Department of Medicine, Uni-versity in Oslo, Oslo, Norway.Received for publication April 12, 2014; accepted June 19, 2014.Reprints: Anne Catrine Trægde Martinsen, PhD, The Intervention Centre,

Oslo University Hospital, Pb 4950 Nydalen, 0424 Oslo, Norway(e‐mail: [email protected]).

The authors declare no conflict of interest.Copyright © 2014 by Lippincott Williams & Wilkins

J Comput Assist Tomogr • Volume 00, Number 00, Month/Month 201

Copyright © 2014 Lippincott Williams & Wilkins. Unau

higher radiation dose. Filtered back-projection (FBP) has, untilrecently, been used as the standard reconstruction method in CT,but all vendors now offer several alternative reconstruction op-tions. Iterative reconstruction (IR) accurately models the datacollection process in CT, and a purely IR technique not onlymodels photon and noise statistics but also includes details withregard to scanner geometry and optics. However, because pureIR requires considerable computer processing power and is timeconsuming, several hybrid solutions combining FBP and IR havebeen developed.2–8

Toshiba's Adaptive Iterative Dose Reduction (AIDR) is ahybrid technique, and the most recent, sophisticated version isreferred to as AIDR 3 dimensional (3D). This iterative algorithmperforms noise reduction in the raw data and image domain. Pro-cessing in the raw data domain involves a statistical model, ascanner model, and projection noise estimation to reduce noise.In the image domain, iterative noise reduction is performed tooptimize the reconstruction.9 Studies have shown a potential fordose reduction using AIDR 3D compared with FBP.10,11

However, uncertainty remains regarding the effect of IRtechniques and reconstruction algorithms on image quality. Awiderange of reconstruction options challenge the pursuit of the opti-mal reconstruction option for each examination.

In the present study, phantom and patient images from theAquilion ONE CT scanner (Toshiba Medical Systems, Tokyo,Japan) were reconstructed with AIDR 3D and FBP. Image qualityfor both techniques was compared using 12 different abdominalreconstruction filters.

Our institution is a referral center for hepatobiliary surgery,and the detection of subcentimeter lesions in the liver parenchymais of utmost importance when evaluating the possibility of surgicaltreatment. Several patients have repeated examinations, increasingthe cumulative radiation dose. The aim of this study was to opti-mize image quality in relation to radiation dose in CT liver exam-inations by comparing images reconstructed with different filterswith and without IR.

MATERIALS AND METHODS

Phantom Study

Image ReconstructionAvariety of reconstruction algorithms are available with the

Toshiba Aquilion ONE CT scanner. Toshiba's family of recon-struction filters based on FBP is named FC, with a subsequentnumber describing the size of the kernel for filtration of theimage. In general, a lower number will produce a smoother imageand a higher number will produce a sharper, edge-enhanced im-age. For abdominal CT examinations, Toshiba recommended fil-ters FC 03, 04, 05, 07, 08, and 09, which include extra beamhardening correction, or their equivalent without extra beam hard-ening correction, FC 13, 14, 15, 17, 18, and 19 (Toshiba AquilionONE operation manual). In addition, AIDR 3Dmay be selected at

4 www.jcat.org 1

thorized reproduction of this article is prohibited.

mailto:[email protected]

www.jcat.org

FIGURE 1. A, A customized adult-sized, anthropomorphic upper-abdomen phantom designed for ROC analysis. B, CT image ofcustomized adult-sized, anthropomorphic upper-abdomen phantom designed for ROC analysis.

TABLE 2. Image Quality Evaluation Criteria for Clinical Images

Reproduction of renal pelvis and perirenal fasciaReproduction of urinary bladder wallReproduction of vessels in the liver

Andersen et al J Comput Assist Tomogr • Volume 00, Number 00, Month/Month 2014

3 different levels, mild, standard, and strong, in combination withthe chosen filter.

In the phantom study, all images were reconstructed withall the filters mentioned above, both with and without AIDR 3Dstandard.

Data CollectionAll scans were performed on a Toshiba Aquilion ONE CT

scanner using our institution's standard liver CT protocol basedon the manufacturer's recommendations. The protocol used thefollowing parameters: tube voltage of 120 kilovolt (peak), detec-tor collimation of 4 cm, pitch of 0.813, and reconstructed slicewidth of 3 mm. The display field of view differed slightly betweenthe dose levels at 35 and 40 cm. Scans were performed at 2 doselevels with a pitch-corrected, weighted CT dose index (CTDIvol)of 10 and 15 mGy. According to the European Guidelines formultislice CT, the reference level for abdominal CT is 15 mGyCTDIvol and the reference level for liver CT is greater than25 mGy CTDIvol.

12 The chosen dose levels for the phantom studyare therefore clinically relevant for abdominal scans.

Receiver Operating Characteristic Phantomand Evaluation

A customized adult-sized, anthropomorphic upper-abdomenphantom (Figs. 1A, B), specially designed for receiver operatingcharacteristic (ROC) analysis of the detectability of liver lesions,was used in the study.13 The phantom included 4 inserts with le-sions ranging from 2 to 7 mm in diameter. All lesions were filledwith water, and the inserts were interchanged and rotated for the2 dose levels, thereby hindering the readers' recognition of thetest pattern and the position of lesions.

Five low-dose images were included to improve randomness.Detectability of 16 small liver lesions was evaluated in 48 imagesby 2 experienced radiologists independently. Thus, a total of768 lesions were evaluated in the study. All images were shownin a randomized and blinded fashion under standardized view-ing conditions on the same picture archiving and communicationsystem workstation with subdued ambient light. Detectability wasrated according to a 5-point scale as shown in Table 1. Receiveroperating characteristic curves were constructed, and test accuracy

TABLE 1. Evaluation Scale for ROC Phantom

1 Low-contrast object definitely not present in sector2 Low-contrast object most likely not present in sector3 Low-contrast object possibly not present in sector4 Low-contrast object most likely present in sector5 Low-contrast object definitely present in sector

2 www.jcat.org


was defined as the area under the curve (AUC). Areas under thecurve were compared using software Analyse-IT for MicrosoftExcel (version 2.00). Analyse-IT uses the Delong Delong Clarke-Pearson method to compare curves.14

Patient Study

Data CollectionThe 2 reconstruction options that achieved the highest scores

for both dose levels in the phantom study were compared in a pa-tient study. Computed tomography liver examinations of 6 pa-tients, 2 women and 4 men, median age of 63.5 (42–72) years,were included. All patients had abdominal cancers or otherabdominal pathologies. Images were reconstructed with recon-struction filters FC17 and FC18 in combination with AIDR 3Dstandard.

Image EvaluationThe liver examinations were evaluated using visual grading

analysis, in which detail visibility was evaluated on a 4-point scaleusing the European Guidelines' quality criteria on image quality inthe CT liver examinations evaluation form shown in Table 2, scaleshown in Table 3. Two experienced radiologists evaluated all im-ages independently and in random order on the same worksta-tion in subdued ambient light. The results were evaluated withthe Student t test.

RESULTS

Phantom StudyThe images reconstructed with filters FC17 and FC18 in

combination with AIDR 3D were the only options among the8 best images at both dose levels (AUC, 0.96 and 0.94 for

Reproduction of the major branches of the abdominal aortaReproduction of the adrenal glandsReproduction of the intrapancreatic part of the common bile ductReproduction of the gallbladder wallReproduction of the pancreatic ductReproduction of the hepatic veinsReproduction of lesions (if any)Count lesions (if any) (write no.)

© 2014 Lippincott Williams & Wilkins


www.jcat.org

TABLE 3. Evaluation Scale for Clinical Images

0 Detail not visible1 Detail visible, but not delineated or sharp2 Detail visible, delineated, but not sharp3 Detail clearly defined and sharply delineated

TABLE 5. AUCs for Both Observers, AIDR 3D and FBP, for AllReconstruction Filters

Observer 1 Observer 2

AIDR 3D FBP AIDR 3D FBP

FC03/13 0.88 0.88 0.89 0.88FC04/14 0.90 0.86 0.88 0.87FC05/15 0.87 0.83 0.87 0.88FC07/17 0.90 0.88 0.90 0.88FC08/18 0.87 0.90 0.90 0.88FC09/19 0.87 0.88 0.87 0.85

Scores for 10 and 15 mGy with and without beam hardening wereplotted together.

J Comput Assist Tomogr • Volume 00, Number 00, Month/Month 2014 Image Quality in Relation to Radiation Dose in CT

15 mGy as well as 0.86 and 0.84 for 10 mGy) (Table 4). Theimages reconstructed with AIDR 3D scored slightly higher thanthe images reconstructed with FBP. The difference was greatestfor the 10-mGy dose level. Adaptive Iterative Dose Reduction3D did not significantly improve lesion detectability for any ofthe filters.

For the 15-mGy dose level, the score was slightly higherwithout beam hardening correction, a trend that was not observedat 10 mGy. There were only minor differences between the recon-struction algorithms with respect to lesion detectability in thephantom study, and most of the differences were not significant.

Receiver operating characteristic curves were also made forthe 2 observers individually. The data from the 2 dose levels re-constructed with and without beam hardening were plotted to-gether in the same curves. Areas under the curve are given inTable 5. Observer 1 scored highest for the images reconstructedwith FC04/14 and FC07/17 with AIDR 3D as well as FC08/18with FBP (AUC, 0.9). Observer 2 scored highest for the imagesreconstructed with FC07/17 and FC08/18 with AIDR 3D (AUC,0.9). The difference between the observers was significant onlyfor FC05/15 with FBP and for FC08/18 with AIDR 3D.

TABLE 4. Area Under the ROC Curve in Descending Orderfor All Phantom Images

15 mGy 10 mGy

Filter Algorithm AUC Filter Algorithm AUC

FC7 AIDR 3D 0.96 FC17 AIDR 3D 0.86FC17 AIDR 3D 0.96 FC8 FBP 0.86FC18 FBP 0.94 FC8 AIDR 3D 0.85FC19 AIDR 3D 0.94 FC14 AIDR 3D 0.85FC14 FBP 0.94 FC15 AIDR 3D 0.84FC13 FBP 0.94 FC18 AIDR 3D 0.84FC19 FBP 0.94 FC4 AIDR 3D 0.84FC18 AIDR 3D 0.94 FC3 AIDR 3D 0.84FC14 AIDR 3D 0.93 FC13 FBP 0.84FC3 AIDR 3D 0.93 FC5 AIDR 3D 0.84FC13 AIDR 3D 0.93 FC7 AIDR 3D 0.84FC4 AIDR 3D 0.93 FC18 FBP 0.84FC7 FBP 0.93 FC17 FBP 0.84FC17 FBP 0.93 FC7 FBP 0.83FC8 FBP 0.93 FC13 AIDR 3D 0.83FC3 FBP 0.93 FC3 FBP 0.83FC5 FBP 0.91 FC19 FBP 0.83FC9 AIDR 3D 0.91 FC4 FBP 0.83FC8 AIDR 3D 0.91 FC9 FBP 0.83FC5 AIDR 3D 0.91 FC9 AIDR 3D 0.82FC15 AIDR 3D 0.90 FC15 FBP 0.82FC15 FBP 0.90 FC14 FBP 0.82FC4 FBP 0.89 FC19 AIDR 3D 0.81FC9 FBP 0.87 FC5 FBP 0.79



Looking at the observers separately, FC07/17 and FC08/18were the reconstruction filters with the highest score for both ob-servers (AUC, 0.9). The combined scores for both observersshowed that FC17 with AIDR 3D and FC18 with AIDR 3D werethe filters with the highest score for both dose levels. On the basisof this, these reconstruction filters were chosen for the study ofpatient images.

Receiver operating characteristic curves are shown inFigures 2A to C, where A and B reflect the score for filtersFC08 and FC18 with and without AIDR 3D for 15 and 10 mGy,respectively. Figure 2C compares the score for the chosen filter,FC18, with AIDR 3D at 10 mGy, with the poorest reconstructioncombination, FC05without AIDR 3D. Receiver operating charac-teristic curves for the individual observers for filter FC17 andFC18 with AIDR 3D are shown in Figures 2D and E. Figure 2Fshows ROC curves for FC18 with FBP and AIDR 3D for 10 and15 mGy.

Patient StudyIn the patient study, there was a small, nonsignificant differ-

ence in score between the images reconstructed with the 2 filters.The images reconstructed with filter FC18 and AIDR 3D scoredslightly better (mean score, 2.60) than the images reconstructedwith filter FC17 and AIDR 3D (mean score, 2.57), although notsignificantly (P = 0.56). The results for each patient and the meanscore are shown in Table 6.

Interobserver DifferencesThe readers' preference differed in the phantom study. Ob-

server 1 generally preferred FC18 (mean score, 2.24 and 2.35for FC17 and FC18, respectively), whereas observer 2 preferredFC17 (mean score, 2.91 and 2.85 for FC17 and FC18, respec-tively). The difference in score between the filters was not sig-nificant for any of the observers (P = 0.2 and 0.3 for observer 1and observer 2, respectively). Individual results are presented inTable 7 and Figures 3 and 4.

DISCUSSIONIn the phantom study, AIDR 3D performed slightly, although

not significantly, better than FBP, at 10 mGy. This effect was notseen at 15 mGy. This may be due to the greater potential for noisereduction and image quality improvement at lower dose levels ver-sus higher dose levels. Comparison of image quality for FC18with AIDR 3D at 10 mGy compared with FBP reconstruction at15 mGy indicated that AIDR 3D did not compensate for the33% dose reduction in this study.

www.jcat.org 3


www.jcat.org

FIGURE 2. A, ROC curves for filters FC08 and FC18 with and without AIDR 3D IR at 15 mGy. B, ROC curves for filters FC08 and FC18with and without AIDR 3D IR at 10 mGy. C, ROC curves for the chosen filter FC18 with AIDR 3D at 10 mGy and the poorest-performingreconstruction combination, FC05 without AIDR 3D. D, ROC curves for filters FC07/17 for 10 and 15 mGy together. AIDR 3D and FBPseparately for both observers. E, ROC curves for filters FC08/18 for 10 and 15 mGy together. AIDR 3D and FBP separately for both observers.F, ROC curves for filter FC18, both observers. AIDR 3D and FBP for 10 and 15 mGy separately.


At 15 mGy, there was a tendency toward improved perfor-mance without beam hardening correction, whereas this trend wasnot observed at 10 mGy. This could be because the reconstruc-tion algorithm with added beam hardening correction will havea larger potential for improvement of image quality when thereis more absorbed radiation and thereby more noise in the image.

Reconstruction filters FC17 with AIDR 3D as well as FC18with AIDR 3D were chosen for clinical comparison because oftheir superior performance at both dose levels in the phantom study.In the patient study, FC18 with AIDR 3D performed slightly better

4 www.jcat.org


than FC17 with AIDR 3D, although not significantly. The interob-server differences seen in the clinical comparison demonstrategreater divergence in the subjective evaluation of image qualityand texture than in the objective evaluation of diagnostic value.This was also shown in the study by Martinsen et al,2 in whichinterobserver differences were found because of subjective pref-erences and assumed familiarity of image texture.

On the basis of the results from the 2 studies, reconstructionfilter FC18 with AIDR 3D was incorporated into the standard pro-tocol for CT liver examinations in clinical practice in our hospital.



www.jcat.org

TABLE 6. Mean Score for All 6 Patients and Mean Score forthe 2 Reconstruction Filters

FC17 FC18

A 2.59 2.72B 2.40 2.27C 2.67 2.61D 2.83 2.83E 2.47 2.59F 2.45 2.55Mean score 2.57 2.60 FIGURE 3. Individual score for all patients with observer 1.

FIGURE 4. Individual score for all patients with observer 2.

J Comput Assist Tomogr • Volume 00, Number 00, Month/Month 2014 Image Quality in Relation to Radiation Dose in CT

Our study had limitations. The lesions in the phantom studycontained water and may have been easier to detect than a typicalmalignant lesion in a patient. This may partly explain why theAIDR 3D reconstruction was more effective for the images withthe most noise, as many lesions were easy to detect regardless ofreconstruction method for the 15-mGy images. The low rate offalse-positives in the phantom study may be due to somewhat im-proved visibility of lesions compared with a clinical case. Thisaffects the ROC curves resulting in a steep slope to the far left.It seems probable that the phantom study would have shown alarger difference between reconstruction methods if the lesionshad been more difficult to detect. The attenuation difference be-tween lesions and liver parenchyma could have been increasedby using a mixture of water and iodinated contrast or glycerolin the lesions, as was shown by Olerud et al.13 The difference indisplay field of view between the 2 dose levels may affect the de-tectability of lesions in the phantom study. However, this studyaimed to identify differences between reconstruction algorithms.The comparisons of different reconstruction techniques were per-formed per dose level; therefore, differences in field of view be-tween different dose levels will not affect the outcome of thisstudy. In the patient study, only 2 of the patients had liver lesions.In the visual grading analysis, the assumption is made that thevisibility of common abdominal structures reflects the visibilityof lesions. Selection of patients with liver lesions would havebeen beneficial with respect to evaluation of lesion detectability.Variation in patient size and age may also affect image quality,but because all comparisons were performed within patients, thisshould not have affected the study results.

Time constraints and patient accessibility render phantomscanning a suitable choice for protocol development. Collabora-tion between physicists, physicians, and technicians is important.We believe that this method of image optimization is feasiblein clinical practice when a new CT scanner with numerous op-tions for image reconstruction is introduced in a radiologicaldepartment.

TABLE 7. Individual Score for All Patients and Both Observers

Patient

Observer 1 Observer 2

FC17 FC18 FC17 FC18

A 2.13 2.44 3.00 3.00B 2.00 1.75 2.86 2.86C 2.33 2.33 3.00 2.89D 2.67 2.67 3.00 3.00E 2.11 2.33 2.88 2.88F 2.20 2.60 2.70 2.50Mean 2.24 2.35 2.91 2.85



In the jungle of options for image reconstruction on CTscanners today, a systematic approach for optimization is re-quired. This method proved useful and feasible in optimizingimage reconstruction of CT liver examinations on the ToshibaAquilion ONE scanner. The study indicates that the preferredchoice for reconstruction of CT liver examinations performedwith the Toshiba Aquilion ONE should be filter FC18 filter withAIDR 3D IR, although AIDR 3D did not significantly improvelesion detectability.

REFERENCES1. United Nations Scientific Committee on the Effects of Atomic Radiation.

Sources and effects of ionizing radiation. UNSCEAR 2008 Reportto the general assembly with scientific annexes. Vol I, Annex A.United Nations, New York US, 2010. http://www.unscear.org/docs/reports/2008/09-86753_Report_2008_Annex_A.pdf.

2. Martinsen AC, Saether HK, Hol PK, et al. Iterative reconstructionreduces abdominal CT dose. Eur J Radiol. 2012;81:1483–1487.

3. Yasaka K, Katsura M, Akahane M, et al. Model-based iterativereconstruction for reduction of radiation dose in abdominopelvic CT:comparison to adaptive statistical iterative reconstruction. SpringerPlus.2013;2:209.

4. Varutvardhanabhuti V, Sumaira I, Gutteridge C, et al. Comparison ofimage quality between filtered back-projection and the adaptive statisticaland novel model-based iterative reconstruction techniques in abdominalCT for renal calculi. Insights Imaging. 2013;4:661–669.

5. Desai GS, Thabet A, Elias AYA, et al. Comparative assessment of threeimage reconstruction techniques for image quality and radiation dosein patients undergoing abdominopelvic multidetector CT examinations.Br J Radiol. 2013;86:20120161.

6. Kalra MK,Woisetschläger M, DahlströmN, et al. Radiation dose reductionwith sinogram affirmed iterative reconstruction technique for abdominalcomputed tomography. J Comput Assist Tomogr. 2012;36:339–346.

7. Laqmani A, Buhk JH, Henes FO, et al. Impact of a 4th generationiterative reconstruction technique on image quality in low-dosecomputed tomography of the chest in immunocompromised patients.Fortschr Röntgenstr. 2013;185:749–757.

www.jcat.org 5


http://www.unscear.org/docs/reports/2008/09-86753_Report_2008_Annex_A.pdf

http://www.unscear.org/docs/reports/2008/09-86753_Report_2008_Annex_A.pdf

www.jcat.org


8. Ploussi A, Alexopoulou E, Economopoulos N, et al. Patient radiationexposure and image quality evaluation with the use of IDOSE4 iterativereconstruction algorithm in chest-abdomen-pelvis CTexaminations.Radiat Prot Dosimetry. 2014;158:399–405.

9. Joemai R, Veldkamp WJH, Kroft LJM, et al. Adaptive iterative dosereduction 3D versus filtered back projection in CT: evaluationof image quality. AJR Am J Roentgenol. 2013;201:1291–1297.

10. Gervaise A, Osemont B, Louis M, et al. Standard dose versus low-doseabdominal and pelvic CT: comparison between filtered back projectionversus adaptive iterative dose reduction 3D. Diagn Interv Imaging.2014;95:47–53.

6 www.jcat.org


11. Matsuki M, Murakami T, Juri H, et al. Impact of adaptive iterative dosereduction (AIDR) 3D on low-dose abdominal CT: comparison withroutine-dose CTusing filtered back projection. Acta Radiol. 2013;54:869–875.

12. Menzel HG, Schibilla H, Teunen D. eds. European Guidelines on qualitycriteria for Computed Tomography. Luxembourg: EuropeanCommission; 2006.

13. Olerud HM, Olsen J, Skretting A. An anthropomorphic phantom forreceiver operating characteristic studies in CT imaging of liver lesions.Br J Radiol. 1999;72:35–43.

14. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areasunder two or more correlated receiver operating characteristic curves:a nonparametric approach. Biometrics. 1998;44:837–845.



www.jcat.org

Choosing the Best Reconstruction Technique in · PDF fileChoosing the Best Reconstruction...

Documents

Transcript of Choosing the Best Reconstruction Technique in · PDF fileChoosing the Best Reconstruction...