Artifacts from Misaligned CT in Cardiac Perfusion PET/CT Studies: Frequency, Effects, and Potential Solutions

Patient Population

Twenty-eight consecutive patients (20 men, 8 women; age, 63 ± 12 y [mean ± SD]) who were referred to our institution for evaluation of coronary artery disease were enrolled in this study. All patients had been referred for a PET/CT rest/stress perfusion study. Informed patient consent was given.

PET/CT System

Imaging was performed on a Biograph 16 PET/CT (Siemens Medical Solutions), combining a lutetium oxyorthosilicate-based ECAT ACCEL PET with improved detection electronics (Pico3D) and a 16-slice Somatom Emotion CT. The PET scanner has an axial field of view of 16.2 cm and has no septa, so that all PET data was acquired in 3-dimensional (3D) mode (8,9).

Protocol

Transmission data for the thorax were acquired with a low-dose CT scan (120 kV, 26 mA) performed in shallow breathing. After that, patients received a 300- to 500-MBq injection of ¹³NH₃ synchronized with the start of the PET acquisition. The PET rest examination lasted for 10 min, and 30 min later pharmacologically induced stress was achieved by a 6-min infusion of adenosine at 0.16 mg/min/kg of body weight. Patients received a second 300- to 500-MBq ¹³NH₃ injection, and the PET stress examination was acquired for 10 min. Image data from 5 to 10 min after injection were summed and used for further analysis. Both CT and PET were acquired in the arms-down position to improve patient comfort and reduce the probability of motion.

Data Processing

The acquired PET data was reconstructed using the ordered-subsets expectation-maximization (OSEM) algorithm with 4 subsets and 8 iterations. The attenuation map, image containing the attenuation factor for each voxel that is used for attenuation correction, was obtained from the CT image after the necessary transformations, including scaling from x-ray to γ-ray energy and smoothing to match PET spatial resolution (10).

To investigate whether emission–transmission misregistration was at the origin of perfusion defects, we sought to remove the misregistration by realigning the CT to the PET and repeating the PET reconstruction with the aligned CT-based attenuation map. For this purpose, a registration program was developed using IDL (Interactive Data Language; RSI Inc.), allowing 3 different possibilities to realign the PET and CT examinations: manual registration, automatic registration, and an “emission-driven” in-house–developed correction method to modify the heart outline based on the PET data. Details on each of the realignment techniques are provided in the following sections.

For each PET examination acquired, all 3 realignment techniques were applied separately. PET rest and stress scans coming from the same patient were processed independently, as motion could potentially happen in only 1 of both scans. Using each of the realigned CT-based attenuation maps, the PET raw data were reconstructed again. The tracer uptake was quantified before and after realignment by spatially sampling the left ventricle (LV) and projecting the measured activity on a polar map basis (11). The polar map was then divided in 17 segments according to the AHA17 model (12), normalized, and compared with a normal ¹³NH₃ perfusion map. Segments where the uptake differed by >2.5 SD were considered as perfusion defects.

PET/CT Registration

In both manual and automatic PET/CT registration, motion between the PET and CT scans was approximated as being rigid and with no rotational component—that is, registration was limited to a translation between both datasets in all 3 spatial directions.

Manual registration was done by interactively moving the CT image over the PET image and assessing the overlap in fused PET/CT images from coronal, sagittal, and transaxial views. Careful manual registration required 30–60 s for a PET/CT dataset.

We implemented the automatic registration using normalized mutual information as the similarity measure (13)—that is, as the indicator of the similarity between both images in the evaluated position. The principle of automatic image registration is to find the spatial transformation between both images providing the highest value for the similarity measure, and it is commonly achieved through the maximization of the similarity measure using an optimization algorithm. However, the most simple and robust algorithm—although the one that is by far the most computationally expensive—is to do an exhaustive search throughout all transformation parameters, checking all possible spatial transformations to find the one providing the highest similarity measure. The advantage of such an approach is that the algorithm will not stop in the presence of local maxima, so that the absolute maximum of the function is going to be found in all cases. As our transformation was approximated to be purely translational, and with the additional assumption that the displacement in each direction could be no larger than 3 cm, registration using an exhaustive search required a run time of approximately 20 s for each PET/CT dataset.

Emission-Driven Correction

The emission-driven correction is an in-house–developed method based on the following assumption: If there is tracer uptake corresponding to the LV on the PET image, the corresponding voxel in the CT should contain cardiac tissue as well. However, in case there is an inconsistency and the voxel contains lung tissue and, therefore, nearly no attenuation, the value of the voxel is modified to match that of cardiac tissue.

As this operation was only to be applied in the LV, a fully automatic segmentation of the LV from the PET scan was required. The LV always has high ¹³NH₃ uptake, but a simple histogram thresholding for the whole volume would have failed as other organs in the same bed position—for example, the liver–may show higher uptake (Fig. 1A). The assumption that the heart is located on the left side of the patient made it possible to discard half of the image. Furthermore, an a priori knowledge of the acquisition was used: The bed position used for cardiac imaging was always selected so that the LV is located in the central transaxial slice, corresponding to plane 23 of 47 in our scanner. A strong smoothing filter (boxcar average, 90 mm) was applied to this plane (Fig. 1B), so that the maximum value contained in it after the application of the filter was roughly the spatial center of the LV. A 3D ellipsoid approximating the shape of the LV (but slightly larger) was automatically defined around this point, and a binary mask was used to exclude all voxels outside of this ellipsoid (Fig. 1C). As the location of the LV was now defined, a local histogram thresholding within this area was able to classify the voxels as belonging or not belonging to the LV (Fig. 1D).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1. 

Segmentation of LV used by emission-driven algorithm. (A) Central transaxial plane of initial PET image. (B) Image after masking left side and applying a strong smoothing filter; spatial center of the LV is assumed to be roughly the point with maximum intensity in image. (C) 3D ellipsoid defined around spatial center; 3D ellipsoid is a circle in each transaxial plane. (D) Segmentation results after histogram thresholding within ellipsoid.

After the segmentation, the proposed correction was as follows: If a voxel was classified as belonging to the LV on the PET image, we tested whether the CT-based attenuation factor corresponded to heart tissue (i.e., if it was higher than 0.095 cm⁻¹, assuming 0.1 cm⁻¹ is the reference cardiac tissue attenuation factor). If the attenuation factor was lower, it was replaced by the average cardiac attenuation factor from all other voxels classified as belonging to the LV.

The segmentation and modification of the attenuation map required <1-s run time using a standard personal computer. Although the focus of this study was rest/stress ¹³NH₃ perfusion examinations, the emission-driven correction could be applied to most tracers used in cardiac PET/CT, such as ¹⁸FDG and ⁸²Rb, with the only condition that the LV can be segmented out of the PET data.

Manual Registration

The misalignment between the PET and CT datasets was quantified after manual registration using the Euclidian distance between the original and the registered pose—that is, as the square root of the sum of the misalignments in each individual direction. The misalignment averaged 6.1 ± 7.0 mm for the 28 rest acquisitions and 6.0 ± 5.6 mm for the 28 stress acquisitions, revealing that the extent of observed misalignment is not increased despite the pharmacologic stress and the delay in the acquisition.

The spatial distribution of the motion for all 56 examinations was as follows: left–right, 1.3 ± 2.2 mm (range, 0–5.1 mm); anterior–posterior, 1.6 ± 2.9 mm (range, 0–15.4 mm); head–feet, 4.7 ± 6.1 mm (range, 0–23.6 mm). Head–feet motion represented the major component of the misalignment, in agreement with the main direction of the breathing motion. The maximum value of the misalignment for an examination was 29 mm.

Figure 2 shows the distribution of the misalignment according to its magnitude, as well as the changes in measured tracer uptake and defect size observed after reconstruction with the manually registered CT for attenuation correction. The registration had noticeable effects when the misalignment was >6 mm, which occurred in half of the examinations. When the misalignment was >8 mm (29% of the scans), the effects on the tracer uptake were severe, resulting in notable changes of the defect size (average, 6.7% LV) assessed by comparison with a normal ¹³NH₃ database.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2. 

Classification of 56 scans according to degree of misalignment. For each group, 2 parameters indicating effects resulting from manual registration are shown: the average number of segments from the AHA17 model that incurred an uptake change of >10% and the change in defect size in percentage of myocardium.

The mean local differences of normalized uptake between the original studies and the studies after manual registration are shown in Figure 3A. Significant differences were observed in the anterior and anterolateral segments, suggesting that a bias due to misregistration could exist in PET/CT perfusion studies, showing up as slightly reduced uptake in these segments.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 3. 

Mean differences in normalized uptake: (A) between studies with and without manual registration and (B) between studies with manual registration and with emission-driven correction. *P < 0.10; **P < 0.05.

The defect size of 6 of 28 patients (21.4%) changed by >10% LV in either the rest or the stress examinations, all of them after a manual realignment of >10 mm. Of these 6 patients, 4 had a defect (15%–46% LV) located in the anterior or anterolateral segment, which was completely artifactual and disappeared after reconstruction with the realigned CT, 1 patient had an artifactual defect (15% LV) in the inferoapical wall, and another patient had an increase in size (by 13% LV) of a septal perfusion defect induced by the registration. This was the only case where the correction resulted in a modest increase of the defect size.

Automatic Registration

Two different approaches were investigated: When the registration was applied to the complete dataset, the initial PET/CT pose was considered to be optimal by the registration algorithm, as the voxels from the body in the PET and CT images were sufficiently aligned. In other words, the registration algorithm does not improve the cardiac registration when applied to the complete dataset, as the effects of a misaligned heart are technically compensated by a reasonably well-aligned thorax. Subsequently, we tested a more regional approach focusing on the cardiac structures only by manually defining a volume of interest around the heart and removing all other surrounding structures. Unfortunately, this approach also did not produce the desired result, as the low correlation between the functional (only LV) and anatomic images (atria, ventricles, blood) in the heart made mutual information incapable of properly assessing the similarity between both datasets in the evaluated positions.

Emission-Driven Algorithm

Figure 4 shows the comparison between the variation of measured uptake produced by the emission-driven algorithm and by manual registration. Because of the reduced size of the regions considered in the AHA17 model, some local differences were observed. The correlation between both methods was high (R² = 0.74, P < 0.001), and good agreement was seen for large misregistration-induced defects. Mean local differences in uptake between the studies corrected by manual registration and by the emission-driven algorithm were all nonsignificant and were within a 2% interval (Fig. 3B), indicating good agreement between both approaches. Figure 5 shows an example of the emission-driven correction compared with manual registration for a patient with a severe misalignment.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 4. 

Comparison between correction methods. (A) Regression plot between variation of local measured uptake (in percentage of original uptake) introduced by manual realignment (Man Reg) and by emission-driven (Em Dr) algorithm. Each point corresponds to 1 segment of the AHA17 model from perfusion examination of 1 patient. Regression line (solid line) and line of equality (dotted line) are shown. (B) Difference in correction plotted against their average (Bland–Altman plot), with mean difference (dashed line) and 1.96-SD intervals (dotted line).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 5. 

Study with PET/CT misregistration of 11 mm. Columns from left to right: PET, CT, fused PET/CT, and polar map. From top to bottom: (A) Original study. (B) Study after manual registration and reconstruction using coregistered attenuation map. (C) Study after emission-driven correction and reconstruction using modified attenuation map. Undercorrection for attenuation in anteroapical and anterolateral segments produced reduced apparent uptake and was assessed as a perfusion defect (24% of myocardium) after comparison with a normal database. Both correction methods had comparable results, recovering uptake in these segments and demonstrating that defect was completely artifactual.

One theoretic limitation of the emission-driven algorithm—although not observed in this study—is as follows: For a severe perfusion defect located in a misaligned area, the uptake would probably be so low that the segmentation would not identify it as cardiac tissue and would, therefore, not perform any correction for this area—that is, the emission-driven algorithm would preserve this area as it initially was in the original study, where the misregistered perfusion defect would likely appear more intense than it really is. This limitation is related to the segmentation step of the algorithm, which is currently based on thresholding of the image intensity. Improvements of the segmentation step by spatial constraints, model-based segmentation, or use of the tracer dynamics of the myocardium could help overcome it.