- Email: [email protected]
Automated quality control of emission-transmission misalignment for attenuation correction in myocardial perfusion imaging with SPECT-CT systems
Automated quality control of emission-transmission misalignment for attenuation correction in myocardial perfusion imaging with SPECT-CT systems Ji Chen, PhD,a Serpil F. Caputlu-Wilson, MS,b Hongcheng Shi, MD, PhD,ac James R. Galt, PhD,a Tracy L. Faber, PhD,a and Ernest V. Garcia, PhDa Background. Emission-transmission misalignment with single-photon emission computed tomography (SPECT)– computed tomography (CT) systems can impair attenuation correction (AC) in myocardial perfusion imaging. This study was performed to develop automated quality control (Auto-QC) to detect critical misalignment that can significantly impact AC. Methods and Results. Auto-QC was developed to segment myocardium and mediastinum from emission and transmission reconstructions, respectively. Myocardium-mediastinum mismatch was used as the quality-control index (QCI). The QCI threshold for acceptable AC was determined with NCAT (NURBS [nonuniform rational B-spline]– based cardiac torso phantom) simulation and verified with 2 patients with minimal misalignment. Compromised data sets, generated by shifting the attenuation maps by 0.5, 1.0, 1.5, and 2.0 pixels along left-right, up-down, and head-foot directions, respectively, were qualitatively and quantitatively compared with the unshifted data sets. Auto-QC was tested with the 2 verification patients and 41 additional patients. Shifts by more than 1 pixel along any direction compromised AC. Auto-QC with the QCI threshold (3%) had highly concordant results with manual quality control in the detection of critical misalignment (sensitivity of 88% and 90% and specificity of 93% and 95% for the tests by use of the 2 verification patients and 41 additional patients, respectively). Conclusion. QCI quantitatively represented the severity of misalignment. Auto-QC can help clinicians be aware of critical misalignment and can assist in realignment of SPECT and CT images. (J Nucl Cardiol 2006;13:43-9.) Key Words: Attenuation correction • automated quality control • emission-transmission misalignment • myocardial perfusion imaging Attenuation correction (AC) has undergone extensive clinical investigation1-4 and now is a recommended technique for improving the image quality of myocardial perfusion imaging (MPI) with single photon emission computed tomography (SPECT).5 Given that the attenuating material in a patient’s thorax is too varied to meet the constant attenuation coefficient approximation made in the methods of both Sorenson6 and Chang,7 transmission imaging is required to obtain patient-specific attenuation maps for accurate AC in MPI.8-11 It has been From Emory University School of Medicinea and Georgia Institute of Technology, Atlanta, Ga,b and Zhongshan Hospital, Fudan University, Shanghai, China.c Received for publication May 30, 2005; final revision accepted Sept 2, 2005. Reprint requests: Ji Chen, PhD, Assistant Professor of Radiology, Department of Radiology, Emory University, 1364 Clifton Rd, Atlanta, GA 30322; [email protected]. 1071-3581/$32.00 Copyright © 2006 by the American Society of Nuclear Cardiology. doi:10.1016/j.nuclcard.2005.11.007
pointed out that high-quality transmission scans and sufficient transmission counts with low cross-talk from the emission radionuclide are essential to reduce the propagation of noise and error into the AC images.5 Recently, hybrid x-ray computed tomography (CT) and SPECT systems became available for MPI, and AC with these systems has been shown to improve sensitivity, specificity, and predictive accuracy in the detection of coronary artery disease.12 Although there are SPECT-CT systems that can acquire CT and SPECT images simultaneously,13 the available commercial SPECT-CT systems are designed for sequential SPECT and CT imaging. Whereas attenuation maps with high resolution and contrast can be obtained with x-ray CT without the cross-talk and low-count issues that are associated with conventional radionuclide transmission scans, registration of the emission SPECT images and transmission CT images becomes the major qualitycontrol (QC) factor. A recent SPECT-CT study reported that misalignment between the emission and transmis43
44
Chen et al Automated quality control of emission-transmission misalignment
sion scans introduced artifacts at the apical, anterior, and septal wall and proposed a semiautomatic method to detect and correct for SPECT-CT emission-transmission misalignment based on estimation of the body contour from emission and transmission sagittal images by histogram plotting and thresholding.14 It should be noted that this method may not be applicable to large patients with truncated attenuation maps or emission reconstructions. Automatic registration of SPECT and CT images has also been proposed and evaluated with a physical torso phantom15; however, this method requires a radionuclide transmission image to be acquired simultaneously with the SPECT emission image, which is not available or practical with current SPECT-CT systems. This article presents an automated quality-control (Auto-QC) algorithm to detect critical emissiontransmission misalignment that may introduce artifacts to the AC images. With this automatic method incorporated into the routine clinical workflow, physicians would be informed if there is critical misalignment between the SPECT and CT reconstructions and would be assisted in realignment of the SPECT and CT images. MATERIALS AND METHODS Auto-QC Algorithm Auto-QC was developed to compare the objects segmented from the SPECT and CT reconstructions by image segmentation techniques. Image segmentation automatically differentiated the lung and soft tissue in the attenuation map by thresholding and segmented the myocardium in the emission reconstruction by region growing.16,17 Region growing was initiated by automatically defining a region in the myocardium. In this region suitably defined properties (count level and gradient) that reflected membership in the myocardium were computed. Next, the boundaries of this region were examined. A given boundary was strong if the properties differed significantly on either side of that boundary, and it was weak if they did not. Strong boundaries were allowed to stand, whereas weak boundaries were dissolved and the adjacent pixels merged. The process was iterated by alternately recomputing the membership properties for the enlarged region and then dissolving weak boundaries. The segmentation was complete when a point was reached at which no boundaries were weak enough to be dissolved. After image segmentation, morphologic processing was applied to the objects obtained from the emission reconstruction (myocardium) and from the attenuation map (soft tissue and lung) to improve the results from image segmentation.18 Morphologic opening and closing were used to reduce the impact on image segmentation from streaking artifacts in the attenuation map as a result of patient respiratory or body motion during the CT scan and from extracardiac activities in the emission reconstruction.
Journal of Nuclear Cardiology January/February 2006
After image segmentation and morphologic processing, three images were obtained: (1) a myocardium image from the emission reconstruction, (2) a lung image from the attenuation map, and (3) a soft-tissue image from the attenuation map. The mediastinum image was then obtained from the lung and soft-tissue images as soft tissue between the two lungs. These segmented object images were 3-dimensional binary images in which a pixel value equal to 1 meant that the pixel was inside the object and a value equal to 0 meant that the pixel was outside the object. Taking a logical operation “AND” between the myocardium and mediastinum images produced an image with the matched myocardium and mediastinum. Qualitycontrol index (QCI) was then calculated as follows: (Total counts of myocardium image ⫺ Total counts of matched myocardium-mediastinum image)/Total counts of myocardium image.
Determination of QCI Threshold The 3-dimensional activity distribution and attenuation map were simulated by use of NCAT (NURBS [nonuniform rational B-spline]– based cardiac torso phantom) to represent a normal subject (Figure 1).19,20 One set of SPECT projections with a count level comparable to typical clinical cases (60 projections, 180° orbit) was generated from the activity and attenuation maps with Poison noise. The simulated attenuation map was then shifted by 0.5, 1.0, 1.5, and 2.0 pixels (6.4 mm/pixel) along x (left-right), y (up-down), and z (head-foot) directions, respectively. AC images were then reconstructed by use of the original attenuation map (baseline) and shifted attenuation maps (compromised with emission-transmission misalignment), respectively, with the same reconstruction, filtering, and reorientation parameters. A blinded nuclear medicine physician compared the compromised AC images with the baseline AC image and divided them into two groups: those with artifacts and those without artifacts. The QCI threshold, above which Auto-QC flagged “critical misalignment” and below which Auto-QC flagged “no/noncritical misalignment,” was then determined so that Auto-QC grouped the simulation studies in the same manner as the physician’s assessment. Two patient studies acquired by a GE Millennium VG/ Hawkeye SPECT-CT system (GE Medical Systems, Milwaukee, Wisc) were used to verify the QCI threshold determined by the simulation study. Subjective assessment of these 2 patients by use of the vendor-supplied, manual-registration QC software (Xeleris 1.1; GE Medical Systems, Milwaukee, Wisc) indicated that there was minimal misalignment between the SPECT and CT images (baseline). Varying degrees of misalignment as was done for the simulation study were introduced. All of the data sets with introduced misalignment were divided into 2 groups, those with critical misalignment and those without critical misalignment, based on the result of the simulation study. Absolute differences in the maximal and mean myocardial counts (XAD and MAD) to baseline were calculated for all of the data sets by region-of-interest analyses and compared between the two groups. Auto-QC was performed on all of the
Journal of Nuclear Cardiology Volume 13, Number 1;43-9
Chen et al Automated quality control of emission-transmission misalignment
45
Figure 2. Transaxial (A), coronal (B), and sagittal (C) displays. Figure 1. Activity map (A), attenuation map (B), simulated SPECT projections (C), and reconstructed transaxial images (D).
data sets, and QCI was calculated and compared between the two groups.
the QCI threshold was performed on each study to indicate that there was no/noncritical or critical misalignment. These flags were then compared with those given by subjective assessment by use of the vendor-supplied software (Xeleris 1.1; GE Medical Systems, Milwaukee, Wisc) by the expert physicist.
RESULTS Testing of QCI Threshold and Auto-QC Algorithm The data sets with varying degrees of misalignment, generated from the 2 verification patients, were evaluated by a blinded nuclear medicine physicist with extensive AC experience. Using the vendor-supplied software (Xeleris 1.1; GE Medical Systems, Milwaukee, Wisc), the physicist marked each data set as no/noncritical or critical misalignment and compared this grouping with the grouping based on the QCI threshold. An additional 41 (24 ⫹ 17) retrospective patient studies acquired by 2 GE Millennium VG/Hawkeye SPECT-CT systems were used to test Auto-QC. Auto-QC with
Figure 2 shows the emission and transmission images before and after image segmentation in one simulation study. Table 1 shows the calculated QCI and visual assessment of the simulation studies without and with shifts of the attenuation map. Artifacts were observed by the blinded physician in the AC images, which were reconstructed with the attenuation maps shifted by more than 1 pixel. Artifacts appeared at different locations for the shifts to different directions. Some examples are shown in Figure 3, and the complete list is given in Table
46
Chen et al Automated quality control of emission-transmission misalignment
Journal of Nuclear Cardiology January/February 2006
Table 1. Visual assessment and calculated QCI of simulation studies with emission-transmission misalignments
Direction None Right
Left
Down
Up
Foot
Head
Shift (pixels)
QCI (%)
0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0
0 0 0 0 0 0.19 1.59 5.17 10.76 0.57 1.85 3.40 5.22 0.21 1.62 3.86 6.11 0.06 0.77 3.07 6.53 0.19 0.70 3.35 5.40
Artifacts
No No LAT LAT No No SEP SEP No No APEX APEX No No INF/BASE INF/BASE No No ANT ANT No No INF/LAT INF/LAT
The attenuation map was shifted to the right, left, down, up, foot, and head directions by 0.5, 1.0, 1.5, and 2.0 pixels (6.4 mm/pixel), respectively. QCI, defined as the percentage of the emission myocardium image misaligned with the transmission mediastinum image, was calculated by Auto-QC. Visual comparison of the AC images obtained with the shifted attenuation maps to the AC images obtained with the unshifted attenuation map indicated whether and where the shifts introduced artifacts to the AC images. LAT, Lateral wall; SEP, septal wall; INF, inferior wall; ANT, anterior wall.
1. QCI increased as the shift of the attenuation map increased, except those to the right. The Auto-QC algorithm did not detect this type of emission-transmission misalignment, because it shifted the myocardium (mainly the left ventricle [LV]) in the emission images into the right ventricle (RV) in the attenuation map. The RV had the same attenuation coefficients as or attenuation coefficients similar to the LV and was within the segmented mediastinum image; thus there were no mismatches associated with this type of misalignment. QCI was greater than 3% for the shifts to the directions other than right that created artifacts in the AC images.
Figure 3. A, AC images with no misalignment (A) and with 2-pixel misalignment in the x direction (B), y direction (C), and z direction (D).
Comparison between the calculated XAD and MAD for the data sets generated from the 2 verification patients showed that the group with critical misalignment (shifts of ⬎1 pixel) had significantly higher XAD and MAD than those in the group with noncritical misalignment (shifts of ⱕ1 pixel). The XAD and MAD were 6.16% and 6.22% versus 14.47% and 14.69% for patient 1 (P ⬍ .0001) and 5.18% and 5.05% versus 12.34% and 12.14% for patient 2 (P ⬍ .0001), respectively, between the noncritical and critical groups. This comparison indicated that significant distortion of the AC images resulted from the critical emission-transmission misalignment. Table 2 summarized the calculated QCI for both groups. The QCI threshold determined by the simulation study (3%) was higher than the maximal QCI in the group with noncritical misalignment and lower than the minimal QCI in the group with critical misalignment, verifying that the QCI threshold determined by the simulation was applicable to clinical patient studies. Auto-QC by use of the determined QCI threshold
Journal of Nuclear Cardiology Volume 13, Number 1;43-9
Chen et al Automated quality control of emission-transmission misalignment
Table 2. Comparison of QCI between groups with noncritical and critical misalignments
Table 3. Testing of QCI threshold and Auto-QC algorithm
QCI (%) Direction Right Left Down Up Foot Head
Manual QC
Shift (pixels)
Patient 1
Patient 2
0.5, 1.5, 0.5, 1.5, 0.5, 1.5, 0.5, 1.5, 0.5, 1.5, 0.5, 1.5,
0.00, 0.00 0.00, 0.00 0.80, 1.63 4.19, 7.98 0.66, 1.65 3.64, 7.07 1.45, 2.87 5.60, 10.46 1.53, 2.72 5.76, 7.94 1.06, 1.52 3.42, 4.56
0.00, 0.00, 0.67, 3.88, 0.35, 3.20, 1.31, 4.64, 1.21, 5.01, 0.98, 3.11,
1.0 2.0 1.0 2.0 1.0 2.0 1.0 2.0 1.0 2.0 1.0 2.0
47
0.00 0.00 1.66 7.01 1.83 5.95 2.21 8.42 2.07 6.83 1.04 4.32
The attenuation maps of the 2 verification patients were shifted to the right, left, down, up, foot, and head directions by 0.5, 1.0, 1.5, and 2.0 pixels (6.4 mm/pixel), respectively. As determined by the simulation study, shifts by more than 1 pixel along any direction impacted AC and were considered as critical misalignment. The calculated QCI values were compared between the two groups: those with noncritical misalignment (shifts of ⱕ 1 pixel) and those with critical misalignment (shifts of ⬎1 pixel).
was tested with all of the data sets generated from the 2 verification patients. When the cases with right shifts, which moved the LV in the emission reconstructions into the RV in the attenuation maps and could not be detected by Auto-QC, were excluded, high concordance (sensitivity of 90% [19/21] and specificity of 95% [18/19]) between Auto-QC and manual QC was achieved (Table 3). This algorithm was also tested in 41 additional patient studies. An example is shown in Figure 4. This patient was positioned with a correct table height for his stress scan but with an incorrect table height for his rest scan. Minor emission-transmission misalignment was detected in the stress study (QCI, 0.60%), and Auto-QC flagged this study as having no/noncritical misalignment. Critical misalignment was detected in the rest scan (QCI, 9.95%) and compromised the AC images. There were decreased counts at the apex (Figure 4B), which should be artifacts due to the misalignment because the apex was aligned with the lung and underwent insufficient AC. Manual realignment of the emission and transmission images of the rest study improved the quality of the AC images and removed the apical artifacts (Figure 4C). The sensitivity and specificity of detection of critical misalignment in these 41 patient studies were 88% (23/26) and 93% (14/15), respectively (Table 3).
Auto-QC Test A* No/noncritical Critical Test B§ No/noncritical Critical
No/noncritical
Critical
22† (18‡) 1† (1‡)
6† (2‡) 19† (19‡)
14 1
3 23
*Test A was the test in which all of the data sets generated from the 2 verification patients were used. There were 48 cases in total, 24 for each patient. Excluding the cases with shifts to the right, the total number of cases was 40, 20 for each patient. † These numbers were obtained from all of the data sets including the cases with shifts to the right, which moved the LV in the emission images into the region of the RV in the attenuation map and were not detected well by Auto-QC. ‡ These numbers were obtained when the cases with right shifts were excluded. § Test B was the test that included the 41 additional patients.
Figure 4. A, The stress study with no/non-critical missalignment. The rest study with critical misalignment (B), and after manual realignment (C).
48
Chen et al Automated quality control of emission-transmission misalignment
DISCUSSION This article studied the impact of SPECT-CT emission-transmission misalignment on AC in MPI with the NCAT phantom. The simulation found that AC by use of the attenuation maps shifted by more than 1 pixel (6.4 mm) created artifacts in the reconstructed emission images. The artifacts appeared at different locations in the myocardium for the shifts to different directions. The impact of critical misalignment on AC has also been shown by use of 2 patients with minimal misalignment. The XAD and MAD were compared between the groups with noncritical (ⱕ1 pixel) and critical (⬎1 pixel) misalignments and showed that significant distortion of the myocardial counts resulted from the critical misalignment. This finding to some extent matched what was found in a previous study of SPECT-CT emission-transmission misalignment with patients, which indicated that a 7-mm shift of the registered attenuation map, regardless of its direction, produced up to a 15% change in the relative regional activity.21 The limitation in the simulation was that Compton scatter and distancedependent collimator resolution were not included in the projector. The simulation should have better image contrast than that in real patients such that it might enable the physician to more easily identify the artifacts in the compromised AC images. Although the QCI threshold determined by the NCAT simulation might not be optimal, it was verified with the 2 patients with minimal misalignment and seemed to be applicable to clinical patient studies. It must be noted that the simulation only included misalignment along a single axis, whereas in practice the SPECT-CT misalignment can be along 2 or 3 axes. This can be another degradation factor of the QCI threshold developed from the simulation study. The developed Auto-QC algorithm automatically segmented the myocardium from the emission image and the mediastinum from the attenuation map and calculated the myocardium-mediastinum mismatch as QCI for identification of critical misalignment. The advantages of this algorithm were that (1) it was totally automatic and (2) its QCI quantitatively represented the severity of emission-transmission misalignment. The major limitation of the Auto-QC algorithm was that it could not detect the misalignment that shifted the LV in the emission image into the RV in the transmission image. This might reduce the accuracy of this algorithm in some studies. Excluding the cases with this type of misalignment, this algorithm yielded highly concordant results with the manual QC by the blinded physicist (sensitivity of 90% and specificity of 95%) in the 40 cases generated from the 2 verification patients. Auto-QC was tested in
Journal of Nuclear Cardiology January/February 2006
41 additional patient studies and had a sensitivity of 88% and specificity of 93% in detection of critical misalignment. In practice most of the cases with significant emission-transmission misalignment have single-axis shifts along the up-down direction or multi-axis shifts including the up-down direction as a result of the so-called diving-board effect. The diving-board phenomenon is usually observed with heavy patients, where the table height is remarkably different between the emission and transmission scans. The tests of the Auto-QC algorithm were based on manual QC by the blinded physicist as the gold standard. Although manual QC should be a routine QC procedure for AC with SPECT-CT systems, its accuracy is limited even with the aid of the vendor-supplied, manualregistration QC software. A prospective trial with a better gold standard should be the next step to validate the Auto-QC algorithm. Positron emission tomography–CT studies could be included in this trial in addition to SPECT-CT studies, because this algorithm is expected to be applicable to positron emission tomography–CT studies. In conclusion, emission-transmission misalignment of more than 1 pixel (6.4 mm/pixel) along any direction can have a negative impact on AC. Identification of the SPECT-CT studies with this amount of emission-transmission misalignment or higher is necessary to judge whether accurate AC can be obtained. The developed Auto-QC algorithm, which can automatically identify critical misalignment and had a sensitivity of 88% and specificity of 93% in a preliminary validation with 41 retrospective patient studies, is designed to help clinicians be aware of critical misalignment as an aid in their decision of whether to perform re-registration of SPECT and CT images before AC in a particular study. Acknowledgment The authors have indicated they have no financial conflicts of interest.
References 1. Hendel RC, Berman DS, Cullom SJ, Follansbee W, Heller GV, Kiat H, et al. Multicenter clinical trial to evaluate the efficacy of correction for photon attenuation and scatter in SPECT myocardial perfusion imaging. Circulation 1999;99:2742-9. 2. Links JM, Becker LC, Rigo P, Taillefer R, Hanelin L, Anstett F, et al. Combined corrections for attenuation, depth-dependent blur, and motion in cardiac SPECT: a multicenter trial. J Nucl Cardiol 2000;7:414-25. 3. Shotwell M, Singh BM, Fortman C, Bauman BD, Lukes J, Gerson MC. Improved coronary disease detection with quantitative attenuation-corrected Tl-201 images. J Nucl Cardiol 2002;9:52-61.
Journal of Nuclear Cardiology Volume 13, Number 1;43-9
4. Links JM, DePuey G, Taillefer R, Becker LB. Attenuation correction and gating synergistically improve the diagnostic accuracy of myocardial perfusion SPECT. J Nucl Cardiol 2002;9:183-7. 5. Hendel RC, Corbett JR, Cullom SJ, DePuey EG, Garcia EV, Bateman TM. The value and practice of attenuation correction for myocardial perfusion SPECT imaging: a joint position statement from the American Society of Nuclear Cardiology and the Society of Nuclear Medicine. J Nucl Cardiol 2002;9:135-43. 6. Sorenson JA. Quantitative measurement of radioactivity in vivo by whole body counting. In: Hine JH, Sorenson JA, editors. Instrumentation in nuclear medicine. New York: Academic Press; 1974. p. 311-48. 7. Chang LT. A method for attenuation correction in radionuclide computed tomography. IEEE Trans Nucl Sci 1978;1:638-43. 8. Tsui BMW, Gullberg GT, Edgerton ER, Ballard JG, Perry JR, McCartney WH, et al. Correction of nonuniform attenuation in cardiac SPECT imaging. J Nucl Med 1989;30:497-507. 9. King MA, Tsui BMW, Pan TS. Attenuation compensation for cardiac single-photon emission computed tomographic imaging: part 1. Impact of attenuation and methods of estimating attenuation maps. J Nucl Cardiol 1995;2:513-24. 10. Zaidi H, Hasegawa B. Determination of the attenuation map in emission tomography. J Nucl Med 2003;44:291-315. 11. American Society of Nuclear Cardiology. Updated imaging guidelines for nuclear cardiology procedures, part 1. J Nucl Cardiol 2001;8:G5-58. 12. Liu Y, Wackers FJ, Natale D, et al. Validation of a hybrid SPECT/CT system with attenuation correction: a phantom study and multicenter trial [abstract]. J Nucl Med 2003;44:290P.
Chen et al Automated quality control of emission-transmission misalignment
49
13. Bocher M, Balan A, Krausz Y, Shrem Y, Lonn A, Wilk M, et al. Gamma camera-mounted anatomical X-ray tomography: technology, system characteristics and first images. Eur J Nucl Med 2000;27:619-27. 14. Fricke H, Fricke E, Weise R, Kammeier A, Lindner O, Burchert W. A method to remove artifacts in attenuation-corrected myocardial perfusion SPECT introduced by misalignment between emission scan and CT-derived attenuation maps. J Nucl Med 2004;45: 1619-25. 15. Dey D, Slomka PJ, Hahn LJ, Kloiber R. Automatic threedimensional multimodality registration using radionuclide transmission CT attenuation maps: a phantom study. J Nucl Med 1999;40:448-55. 16. Zucker S. Region growing: childhood and adolescence. Comput Graph Image Process 1976;5:382-99. 17. Aggarwal JK, Duda RO, Rosenfeld A. Computer methods in image analysis. New York: IEEE Press; 1977. 18. Dougherty ER. An introduction to morphological image processing. Bellingham (WA): SPIE Press; 1992. 19. Segars WP, Lalush DS, Tsui BMW. A realistic spine-based dynamic heart phantom. IEEE Trans Nucl Sci 1999;46:503-6. 20. Segars WP, Tsui BMW, Lalush DS, Frey EC, King MA. Development and application of the new dynamic NURBSbased cardiac-torso (NCAT) phantom [abstract]. J Nucl Med 2001;42:7P. 21. Matsunari I, Boning G, Ziegler SI, Kosa I, Nekolla SG, Ficaro EP, et al. Effects of misalignment between transmission and emission scans on attenuation-corrected cardiac SPECT. J Nucl Med 1998; 39:411-6.
pointed out that high-quality transmission scans and sufficient transmission counts with low cross-talk from the emission radionuclide are essential to reduce the propagation of noise and error into the AC images.5 Recently, hybrid x-ray computed tomography (CT) and SPECT systems became available for MPI, and AC with these systems has been shown to improve sensitivity, specificity, and predictive accuracy in the detection of coronary artery disease.12 Although there are SPECT-CT systems that can acquire CT and SPECT images simultaneously,13 the available commercial SPECT-CT systems are designed for sequential SPECT and CT imaging. Whereas attenuation maps with high resolution and contrast can be obtained with x-ray CT without the cross-talk and low-count issues that are associated with conventional radionuclide transmission scans, registration of the emission SPECT images and transmission CT images becomes the major qualitycontrol (QC) factor. A recent SPECT-CT study reported that misalignment between the emission and transmis43
44
Chen et al Automated quality control of emission-transmission misalignment
sion scans introduced artifacts at the apical, anterior, and septal wall and proposed a semiautomatic method to detect and correct for SPECT-CT emission-transmission misalignment based on estimation of the body contour from emission and transmission sagittal images by histogram plotting and thresholding.14 It should be noted that this method may not be applicable to large patients with truncated attenuation maps or emission reconstructions. Automatic registration of SPECT and CT images has also been proposed and evaluated with a physical torso phantom15; however, this method requires a radionuclide transmission image to be acquired simultaneously with the SPECT emission image, which is not available or practical with current SPECT-CT systems. This article presents an automated quality-control (Auto-QC) algorithm to detect critical emissiontransmission misalignment that may introduce artifacts to the AC images. With this automatic method incorporated into the routine clinical workflow, physicians would be informed if there is critical misalignment between the SPECT and CT reconstructions and would be assisted in realignment of the SPECT and CT images. MATERIALS AND METHODS Auto-QC Algorithm Auto-QC was developed to compare the objects segmented from the SPECT and CT reconstructions by image segmentation techniques. Image segmentation automatically differentiated the lung and soft tissue in the attenuation map by thresholding and segmented the myocardium in the emission reconstruction by region growing.16,17 Region growing was initiated by automatically defining a region in the myocardium. In this region suitably defined properties (count level and gradient) that reflected membership in the myocardium were computed. Next, the boundaries of this region were examined. A given boundary was strong if the properties differed significantly on either side of that boundary, and it was weak if they did not. Strong boundaries were allowed to stand, whereas weak boundaries were dissolved and the adjacent pixels merged. The process was iterated by alternately recomputing the membership properties for the enlarged region and then dissolving weak boundaries. The segmentation was complete when a point was reached at which no boundaries were weak enough to be dissolved. After image segmentation, morphologic processing was applied to the objects obtained from the emission reconstruction (myocardium) and from the attenuation map (soft tissue and lung) to improve the results from image segmentation.18 Morphologic opening and closing were used to reduce the impact on image segmentation from streaking artifacts in the attenuation map as a result of patient respiratory or body motion during the CT scan and from extracardiac activities in the emission reconstruction.
Journal of Nuclear Cardiology January/February 2006
After image segmentation and morphologic processing, three images were obtained: (1) a myocardium image from the emission reconstruction, (2) a lung image from the attenuation map, and (3) a soft-tissue image from the attenuation map. The mediastinum image was then obtained from the lung and soft-tissue images as soft tissue between the two lungs. These segmented object images were 3-dimensional binary images in which a pixel value equal to 1 meant that the pixel was inside the object and a value equal to 0 meant that the pixel was outside the object. Taking a logical operation “AND” between the myocardium and mediastinum images produced an image with the matched myocardium and mediastinum. Qualitycontrol index (QCI) was then calculated as follows: (Total counts of myocardium image ⫺ Total counts of matched myocardium-mediastinum image)/Total counts of myocardium image.
Determination of QCI Threshold The 3-dimensional activity distribution and attenuation map were simulated by use of NCAT (NURBS [nonuniform rational B-spline]– based cardiac torso phantom) to represent a normal subject (Figure 1).19,20 One set of SPECT projections with a count level comparable to typical clinical cases (60 projections, 180° orbit) was generated from the activity and attenuation maps with Poison noise. The simulated attenuation map was then shifted by 0.5, 1.0, 1.5, and 2.0 pixels (6.4 mm/pixel) along x (left-right), y (up-down), and z (head-foot) directions, respectively. AC images were then reconstructed by use of the original attenuation map (baseline) and shifted attenuation maps (compromised with emission-transmission misalignment), respectively, with the same reconstruction, filtering, and reorientation parameters. A blinded nuclear medicine physician compared the compromised AC images with the baseline AC image and divided them into two groups: those with artifacts and those without artifacts. The QCI threshold, above which Auto-QC flagged “critical misalignment” and below which Auto-QC flagged “no/noncritical misalignment,” was then determined so that Auto-QC grouped the simulation studies in the same manner as the physician’s assessment. Two patient studies acquired by a GE Millennium VG/ Hawkeye SPECT-CT system (GE Medical Systems, Milwaukee, Wisc) were used to verify the QCI threshold determined by the simulation study. Subjective assessment of these 2 patients by use of the vendor-supplied, manual-registration QC software (Xeleris 1.1; GE Medical Systems, Milwaukee, Wisc) indicated that there was minimal misalignment between the SPECT and CT images (baseline). Varying degrees of misalignment as was done for the simulation study were introduced. All of the data sets with introduced misalignment were divided into 2 groups, those with critical misalignment and those without critical misalignment, based on the result of the simulation study. Absolute differences in the maximal and mean myocardial counts (XAD and MAD) to baseline were calculated for all of the data sets by region-of-interest analyses and compared between the two groups. Auto-QC was performed on all of the
Journal of Nuclear Cardiology Volume 13, Number 1;43-9
Chen et al Automated quality control of emission-transmission misalignment
45
Figure 2. Transaxial (A), coronal (B), and sagittal (C) displays. Figure 1. Activity map (A), attenuation map (B), simulated SPECT projections (C), and reconstructed transaxial images (D).
data sets, and QCI was calculated and compared between the two groups.
the QCI threshold was performed on each study to indicate that there was no/noncritical or critical misalignment. These flags were then compared with those given by subjective assessment by use of the vendor-supplied software (Xeleris 1.1; GE Medical Systems, Milwaukee, Wisc) by the expert physicist.
RESULTS Testing of QCI Threshold and Auto-QC Algorithm The data sets with varying degrees of misalignment, generated from the 2 verification patients, were evaluated by a blinded nuclear medicine physicist with extensive AC experience. Using the vendor-supplied software (Xeleris 1.1; GE Medical Systems, Milwaukee, Wisc), the physicist marked each data set as no/noncritical or critical misalignment and compared this grouping with the grouping based on the QCI threshold. An additional 41 (24 ⫹ 17) retrospective patient studies acquired by 2 GE Millennium VG/Hawkeye SPECT-CT systems were used to test Auto-QC. Auto-QC with
Figure 2 shows the emission and transmission images before and after image segmentation in one simulation study. Table 1 shows the calculated QCI and visual assessment of the simulation studies without and with shifts of the attenuation map. Artifacts were observed by the blinded physician in the AC images, which were reconstructed with the attenuation maps shifted by more than 1 pixel. Artifacts appeared at different locations for the shifts to different directions. Some examples are shown in Figure 3, and the complete list is given in Table
46
Chen et al Automated quality control of emission-transmission misalignment
Journal of Nuclear Cardiology January/February 2006
Table 1. Visual assessment and calculated QCI of simulation studies with emission-transmission misalignments
Direction None Right
Left
Down
Up
Foot
Head
Shift (pixels)
QCI (%)
0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0
0 0 0 0 0 0.19 1.59 5.17 10.76 0.57 1.85 3.40 5.22 0.21 1.62 3.86 6.11 0.06 0.77 3.07 6.53 0.19 0.70 3.35 5.40
Artifacts
No No LAT LAT No No SEP SEP No No APEX APEX No No INF/BASE INF/BASE No No ANT ANT No No INF/LAT INF/LAT
The attenuation map was shifted to the right, left, down, up, foot, and head directions by 0.5, 1.0, 1.5, and 2.0 pixels (6.4 mm/pixel), respectively. QCI, defined as the percentage of the emission myocardium image misaligned with the transmission mediastinum image, was calculated by Auto-QC. Visual comparison of the AC images obtained with the shifted attenuation maps to the AC images obtained with the unshifted attenuation map indicated whether and where the shifts introduced artifacts to the AC images. LAT, Lateral wall; SEP, septal wall; INF, inferior wall; ANT, anterior wall.
1. QCI increased as the shift of the attenuation map increased, except those to the right. The Auto-QC algorithm did not detect this type of emission-transmission misalignment, because it shifted the myocardium (mainly the left ventricle [LV]) in the emission images into the right ventricle (RV) in the attenuation map. The RV had the same attenuation coefficients as or attenuation coefficients similar to the LV and was within the segmented mediastinum image; thus there were no mismatches associated with this type of misalignment. QCI was greater than 3% for the shifts to the directions other than right that created artifacts in the AC images.
Figure 3. A, AC images with no misalignment (A) and with 2-pixel misalignment in the x direction (B), y direction (C), and z direction (D).
Comparison between the calculated XAD and MAD for the data sets generated from the 2 verification patients showed that the group with critical misalignment (shifts of ⬎1 pixel) had significantly higher XAD and MAD than those in the group with noncritical misalignment (shifts of ⱕ1 pixel). The XAD and MAD were 6.16% and 6.22% versus 14.47% and 14.69% for patient 1 (P ⬍ .0001) and 5.18% and 5.05% versus 12.34% and 12.14% for patient 2 (P ⬍ .0001), respectively, between the noncritical and critical groups. This comparison indicated that significant distortion of the AC images resulted from the critical emission-transmission misalignment. Table 2 summarized the calculated QCI for both groups. The QCI threshold determined by the simulation study (3%) was higher than the maximal QCI in the group with noncritical misalignment and lower than the minimal QCI in the group with critical misalignment, verifying that the QCI threshold determined by the simulation was applicable to clinical patient studies. Auto-QC by use of the determined QCI threshold
Journal of Nuclear Cardiology Volume 13, Number 1;43-9
Chen et al Automated quality control of emission-transmission misalignment
Table 2. Comparison of QCI between groups with noncritical and critical misalignments
Table 3. Testing of QCI threshold and Auto-QC algorithm
QCI (%) Direction Right Left Down Up Foot Head
Manual QC
Shift (pixels)
Patient 1
Patient 2
0.5, 1.5, 0.5, 1.5, 0.5, 1.5, 0.5, 1.5, 0.5, 1.5, 0.5, 1.5,
0.00, 0.00 0.00, 0.00 0.80, 1.63 4.19, 7.98 0.66, 1.65 3.64, 7.07 1.45, 2.87 5.60, 10.46 1.53, 2.72 5.76, 7.94 1.06, 1.52 3.42, 4.56
0.00, 0.00, 0.67, 3.88, 0.35, 3.20, 1.31, 4.64, 1.21, 5.01, 0.98, 3.11,
1.0 2.0 1.0 2.0 1.0 2.0 1.0 2.0 1.0 2.0 1.0 2.0
47
0.00 0.00 1.66 7.01 1.83 5.95 2.21 8.42 2.07 6.83 1.04 4.32
The attenuation maps of the 2 verification patients were shifted to the right, left, down, up, foot, and head directions by 0.5, 1.0, 1.5, and 2.0 pixels (6.4 mm/pixel), respectively. As determined by the simulation study, shifts by more than 1 pixel along any direction impacted AC and were considered as critical misalignment. The calculated QCI values were compared between the two groups: those with noncritical misalignment (shifts of ⱕ 1 pixel) and those with critical misalignment (shifts of ⬎1 pixel).
was tested with all of the data sets generated from the 2 verification patients. When the cases with right shifts, which moved the LV in the emission reconstructions into the RV in the attenuation maps and could not be detected by Auto-QC, were excluded, high concordance (sensitivity of 90% [19/21] and specificity of 95% [18/19]) between Auto-QC and manual QC was achieved (Table 3). This algorithm was also tested in 41 additional patient studies. An example is shown in Figure 4. This patient was positioned with a correct table height for his stress scan but with an incorrect table height for his rest scan. Minor emission-transmission misalignment was detected in the stress study (QCI, 0.60%), and Auto-QC flagged this study as having no/noncritical misalignment. Critical misalignment was detected in the rest scan (QCI, 9.95%) and compromised the AC images. There were decreased counts at the apex (Figure 4B), which should be artifacts due to the misalignment because the apex was aligned with the lung and underwent insufficient AC. Manual realignment of the emission and transmission images of the rest study improved the quality of the AC images and removed the apical artifacts (Figure 4C). The sensitivity and specificity of detection of critical misalignment in these 41 patient studies were 88% (23/26) and 93% (14/15), respectively (Table 3).
Auto-QC Test A* No/noncritical Critical Test B§ No/noncritical Critical
No/noncritical
Critical
22† (18‡) 1† (1‡)
6† (2‡) 19† (19‡)
14 1
3 23
*Test A was the test in which all of the data sets generated from the 2 verification patients were used. There were 48 cases in total, 24 for each patient. Excluding the cases with shifts to the right, the total number of cases was 40, 20 for each patient. † These numbers were obtained from all of the data sets including the cases with shifts to the right, which moved the LV in the emission images into the region of the RV in the attenuation map and were not detected well by Auto-QC. ‡ These numbers were obtained when the cases with right shifts were excluded. § Test B was the test that included the 41 additional patients.
Figure 4. A, The stress study with no/non-critical missalignment. The rest study with critical misalignment (B), and after manual realignment (C).
48
Chen et al Automated quality control of emission-transmission misalignment
DISCUSSION This article studied the impact of SPECT-CT emission-transmission misalignment on AC in MPI with the NCAT phantom. The simulation found that AC by use of the attenuation maps shifted by more than 1 pixel (6.4 mm) created artifacts in the reconstructed emission images. The artifacts appeared at different locations in the myocardium for the shifts to different directions. The impact of critical misalignment on AC has also been shown by use of 2 patients with minimal misalignment. The XAD and MAD were compared between the groups with noncritical (ⱕ1 pixel) and critical (⬎1 pixel) misalignments and showed that significant distortion of the myocardial counts resulted from the critical misalignment. This finding to some extent matched what was found in a previous study of SPECT-CT emission-transmission misalignment with patients, which indicated that a 7-mm shift of the registered attenuation map, regardless of its direction, produced up to a 15% change in the relative regional activity.21 The limitation in the simulation was that Compton scatter and distancedependent collimator resolution were not included in the projector. The simulation should have better image contrast than that in real patients such that it might enable the physician to more easily identify the artifacts in the compromised AC images. Although the QCI threshold determined by the NCAT simulation might not be optimal, it was verified with the 2 patients with minimal misalignment and seemed to be applicable to clinical patient studies. It must be noted that the simulation only included misalignment along a single axis, whereas in practice the SPECT-CT misalignment can be along 2 or 3 axes. This can be another degradation factor of the QCI threshold developed from the simulation study. The developed Auto-QC algorithm automatically segmented the myocardium from the emission image and the mediastinum from the attenuation map and calculated the myocardium-mediastinum mismatch as QCI for identification of critical misalignment. The advantages of this algorithm were that (1) it was totally automatic and (2) its QCI quantitatively represented the severity of emission-transmission misalignment. The major limitation of the Auto-QC algorithm was that it could not detect the misalignment that shifted the LV in the emission image into the RV in the transmission image. This might reduce the accuracy of this algorithm in some studies. Excluding the cases with this type of misalignment, this algorithm yielded highly concordant results with the manual QC by the blinded physicist (sensitivity of 90% and specificity of 95%) in the 40 cases generated from the 2 verification patients. Auto-QC was tested in
Journal of Nuclear Cardiology January/February 2006
41 additional patient studies and had a sensitivity of 88% and specificity of 93% in detection of critical misalignment. In practice most of the cases with significant emission-transmission misalignment have single-axis shifts along the up-down direction or multi-axis shifts including the up-down direction as a result of the so-called diving-board effect. The diving-board phenomenon is usually observed with heavy patients, where the table height is remarkably different between the emission and transmission scans. The tests of the Auto-QC algorithm were based on manual QC by the blinded physicist as the gold standard. Although manual QC should be a routine QC procedure for AC with SPECT-CT systems, its accuracy is limited even with the aid of the vendor-supplied, manualregistration QC software. A prospective trial with a better gold standard should be the next step to validate the Auto-QC algorithm. Positron emission tomography–CT studies could be included in this trial in addition to SPECT-CT studies, because this algorithm is expected to be applicable to positron emission tomography–CT studies. In conclusion, emission-transmission misalignment of more than 1 pixel (6.4 mm/pixel) along any direction can have a negative impact on AC. Identification of the SPECT-CT studies with this amount of emission-transmission misalignment or higher is necessary to judge whether accurate AC can be obtained. The developed Auto-QC algorithm, which can automatically identify critical misalignment and had a sensitivity of 88% and specificity of 93% in a preliminary validation with 41 retrospective patient studies, is designed to help clinicians be aware of critical misalignment as an aid in their decision of whether to perform re-registration of SPECT and CT images before AC in a particular study. Acknowledgment The authors have indicated they have no financial conflicts of interest.
References 1. Hendel RC, Berman DS, Cullom SJ, Follansbee W, Heller GV, Kiat H, et al. Multicenter clinical trial to evaluate the efficacy of correction for photon attenuation and scatter in SPECT myocardial perfusion imaging. Circulation 1999;99:2742-9. 2. Links JM, Becker LC, Rigo P, Taillefer R, Hanelin L, Anstett F, et al. Combined corrections for attenuation, depth-dependent blur, and motion in cardiac SPECT: a multicenter trial. J Nucl Cardiol 2000;7:414-25. 3. Shotwell M, Singh BM, Fortman C, Bauman BD, Lukes J, Gerson MC. Improved coronary disease detection with quantitative attenuation-corrected Tl-201 images. J Nucl Cardiol 2002;9:52-61.
Journal of Nuclear Cardiology Volume 13, Number 1;43-9
4. Links JM, DePuey G, Taillefer R, Becker LB. Attenuation correction and gating synergistically improve the diagnostic accuracy of myocardial perfusion SPECT. J Nucl Cardiol 2002;9:183-7. 5. Hendel RC, Corbett JR, Cullom SJ, DePuey EG, Garcia EV, Bateman TM. The value and practice of attenuation correction for myocardial perfusion SPECT imaging: a joint position statement from the American Society of Nuclear Cardiology and the Society of Nuclear Medicine. J Nucl Cardiol 2002;9:135-43. 6. Sorenson JA. Quantitative measurement of radioactivity in vivo by whole body counting. In: Hine JH, Sorenson JA, editors. Instrumentation in nuclear medicine. New York: Academic Press; 1974. p. 311-48. 7. Chang LT. A method for attenuation correction in radionuclide computed tomography. IEEE Trans Nucl Sci 1978;1:638-43. 8. Tsui BMW, Gullberg GT, Edgerton ER, Ballard JG, Perry JR, McCartney WH, et al. Correction of nonuniform attenuation in cardiac SPECT imaging. J Nucl Med 1989;30:497-507. 9. King MA, Tsui BMW, Pan TS. Attenuation compensation for cardiac single-photon emission computed tomographic imaging: part 1. Impact of attenuation and methods of estimating attenuation maps. J Nucl Cardiol 1995;2:513-24. 10. Zaidi H, Hasegawa B. Determination of the attenuation map in emission tomography. J Nucl Med 2003;44:291-315. 11. American Society of Nuclear Cardiology. Updated imaging guidelines for nuclear cardiology procedures, part 1. J Nucl Cardiol 2001;8:G5-58. 12. Liu Y, Wackers FJ, Natale D, et al. Validation of a hybrid SPECT/CT system with attenuation correction: a phantom study and multicenter trial [abstract]. J Nucl Med 2003;44:290P.
Chen et al Automated quality control of emission-transmission misalignment
49
13. Bocher M, Balan A, Krausz Y, Shrem Y, Lonn A, Wilk M, et al. Gamma camera-mounted anatomical X-ray tomography: technology, system characteristics and first images. Eur J Nucl Med 2000;27:619-27. 14. Fricke H, Fricke E, Weise R, Kammeier A, Lindner O, Burchert W. A method to remove artifacts in attenuation-corrected myocardial perfusion SPECT introduced by misalignment between emission scan and CT-derived attenuation maps. J Nucl Med 2004;45: 1619-25. 15. Dey D, Slomka PJ, Hahn LJ, Kloiber R. Automatic threedimensional multimodality registration using radionuclide transmission CT attenuation maps: a phantom study. J Nucl Med 1999;40:448-55. 16. Zucker S. Region growing: childhood and adolescence. Comput Graph Image Process 1976;5:382-99. 17. Aggarwal JK, Duda RO, Rosenfeld A. Computer methods in image analysis. New York: IEEE Press; 1977. 18. Dougherty ER. An introduction to morphological image processing. Bellingham (WA): SPIE Press; 1992. 19. Segars WP, Lalush DS, Tsui BMW. A realistic spine-based dynamic heart phantom. IEEE Trans Nucl Sci 1999;46:503-6. 20. Segars WP, Tsui BMW, Lalush DS, Frey EC, King MA. Development and application of the new dynamic NURBSbased cardiac-torso (NCAT) phantom [abstract]. J Nucl Med 2001;42:7P. 21. Matsunari I, Boning G, Ziegler SI, Kosa I, Nekolla SG, Ficaro EP, et al. Effects of misalignment between transmission and emission scans on attenuation-corrected cardiac SPECT. J Nucl Med 1998; 39:411-6.