Evaluation of Image Registration in PET/CT of the Liver and Recommendations for Optimized Imaging

Image Acquisition

Hybrid PET/CT scans were acquired using a Biograph Duo (Siemens Medical Solutions USA, Inc.) containing a 2-slice CT scanner. A low-dose CT scan for localization and attenuation-correction purposes was acquired in the caudocranial direction from the thighs to the base of the skull. Scanning parameters included 40 mA·s, 130 kV, 5-mm slice collimation, 0.8-s rotation time, and pitch of 1.5, reconstructed to 3-mm slices for smooth coronal representation. CT scans were acquired during timed unforced expiration breath-hold (EB) or during free breathing (FB). Timed expiration breath-hold consisted of free breathing during the first (caudal) part of the scan, a deep inspiration command at the level of the spina iliaca superior, immediately followed by a command to expire and breath-hold; patients were allowed to resume free breathing at the level of the lung tops. The total expiration breath-hold time was about 30 s. Free breathing was performed without specific patient instructions. No intravenous contrast was applied. For PET, a 3-dimensional (3D) emission scan of the central body was acquired during free breathing, 60 min after intravenous injection of 250 MBq ¹⁸F-FDG. The acquisition time per bed position was 4 min for emission only. Uncorrected emission images as well as images with CT-based attenuation correction were reconstructed, both using 2 iterations, 8 subsets, and a 5-mm 3D gaussian filter. Dedicated ¹⁸F-FDG PET scans were acquired using an ECAT Exact 47 scanner (Siemens Medical Solutions). A 3D emission scan was acquired and reconstructed identical to PET from PET/CT. In addition, a 2-dimensional ⁶⁸Ge-based transmission scan was acquired for attenuation correction. The acquisition time per bed position was 5 min for emission and 3 min for the transmission. Dedicated CT scans were acquired using a Somatom Volume Zoom (Siemens Medical Solutions) 4-slice scanner. Scans of the liver were acquired with 80 mAs, 130 kV, 0.5-s rotation time, and 5-mm slice thickness, during unforced expiration breath-hold. Intravenous contrast was applied; the portal-phase images were selected for image fusion with PET.

Image Registration Procedure

For hybrid PET/CT, normal image registration quality-assurance procedures were followed as described by the manufacturer. This involved alignment of the PET and CT gantries after maintenance, using a “crossed-lines” phantom. No additional image registration optimization was performed after scanning. Software image registration was performed on a personal computer with image viewing and registration software, developed in-house, based on the visualization toolkit VTK (14) and the insight segmentation and registration toolkit ITK (15). The procedure has been described in more detail previously (16). In brief, the software allows rigid-body image registration based on 3 translation and 3 rotation parameters. Anatomic registration of PET emission images to CT was pursued using an implementation of the automatic mutual information algorithm, on a 3D volume of interest containing the liver.

Image Analysis

Image sets from PET and CT were correlated through evaluation of borders of the liver and focal lesions within the liver. Mismatches of >10 mm were considered potentially clinically relevant. Mismatches of focal lesions were expressed as 3D vectors. For liver borders, this approach is not possible, because a unidirectional shift of a liver border may be complicated by a (unrecognizable) deformation or rotation that alters the location that represents the top. Selected for landmarks were the tangent points (tops) of 3 liver borders: the diaphragmatic dome, the right liver border, and the caudal tip. 3D ellipsoids were manually positioned to match the curved shapes of the liver borders (Fig. 1); the locations of the tangent points were then derived mathematically. Mismatches were expressed as 1-dimensional distances along the axis of the largest movement (e.g., the craniocaudal direction for the diaphragmatic dome and the caudal tip of the liver; the lateral direction for the right lateral liver border). This procedure was performed separately on CT, uncorrected (uPET), and attenuation-corrected (acPET) images, blinded from each other.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1. 

Localization of liver borders. Coronal slices of CT (A), attenuation-corrected PET (B), and uncorrected PET (C) of a single patient, acquired with hybrid PET/CT during expiration breath-hold. Circles represent slices through 3D ellipsoids that were mapped to diaphragmatic dome (green), right lateral border (blue), and caudal tip (red) to determine differences in their respective positions.

The localization of liver borders is difficult on uPET and acPET, as the images are blurry. The selected visual cutoff for positioning of a border may be different for uPET and acPET images. The observer-specific systematic bias between localization of liver borders on uPET and acPET was determined by comparing images from dedicated PET, where the position of the liver is theoretically identical on both image sets. The true position of the liver border was assumed to be between the visual localizations on uPET and acPET. All uPET and acPET measurements were corrected afterward for this bias, using the average measurement difference from the theoretic position. The interobserver variability for manual determination of positional differences of tangent points, after correction of the systematic bias, was evaluated in 5 subsequent dedicated PET scans (both uPET and acPET) by 2 experienced observers.

Registration Errors of Liver Borders

The average absolute interobserver variability in determination of liver border position differences on uPET and acPET, measured in 1 direction, was 2 mm (range, −3 to 4 mm; SD, 3 mm) for the diaphragmatic dome, 2 mm (range, −3 to 2 mm; SD, 2 mm) for the right lateral border, and 2 mm (range, −4 to 3 mm; SD, 3 mm) for the caudal tip.

For EB_all, the average absolute image registration error at the diaphragmatic dome of the liver in the craniocaudal direction was 11 mm (relative range, −16 to +44 mm in 40% of cases > 10 mm). Visual inspection revealed that the largest errors were caused by expiration during breath-hold CT that was not deep enough. For the caudal tip of the liver, the average error was 19 mm (range, 0–53 mm in 55% of cases > 10 mm). Registration errors were all <10 mm at the right lateral liver border and were <10 mm at all liver borders in the FB and SF protocols. The image registration errors of FB and SF at the locations of the diaphragmatic dome and caudal tip were significantly less than those of EB_all (P < 0.05). SF did not perform significantly better than FB at the location of all liver borders. The results are listed in more detail in Table 1. The distribution of registration errors peracquisition protocol is represented in Figure 2.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2. 

Image registration errors of liver borders. Relative image registration errors at location of several liver borders for EB (hybrid PET/CT with breath-hold CT), FB (hybrid PET/CT with free-breathing CT), and SF (software fusion of dedicated PET and CT). Registration errors occur primarily in craniocaudal direction (diaphragm and caudal tip affected most) because of insufficient expiration during CT.

The image registration at the diaphragmatic dome in breath-hold PET/CT was not significantly influenced by the adequacy of the breath-hold instructions during CT (error > 10 mm in 44% of EB_adequate and in 36% of EB_failed; not significant). Conversely, the registration of the caudal tip of the liver appeared to be influenced by the success of the breath-hold procedure: error > 10 mm in 33% of EB_adequate and in 73% of EB_failed. However, due to the sample size, this difference did not reach statistical significance. When EB_all was limited to EB_adequate, there was no difference with FB in the diaphragmatic dome. This illustrates that using a faster CT scanner (with more likely successful completion of the breath-hold instructions) does not improve image registration in general, except in the region of the caudal tip where instructions tend to be executed when using a slow CT scanner.

Registration Errors of Liver Lesions

The average interobserver variability in localization of focal liver lesions, measured as a 3D vector, was 2 mm (range, 1–3 mm) on CT and 1 mm (range, 0–2 mm) on acPET.

For EB_all, the average displacement of 11 detected lesions was 11 mm (range, 3–24 mm), with 5 lesions (45%) being displaced >10 mm. There were insufficient evaluable lesions for separate analysis of EB_adequate and EB_failed. For SF, the average displacement of 5 detected lesions was 9 mm (range, 7–14 mm), with 1 lesion displaced >10 mm. Because of the limited number of evaluable lesions, statistical comparison of EB and SF was not performed. An example of a displaced lesion on hybrid PET/CT during EB is shown in Figure 3.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 3. 

Misregistration of a liver lesion on breath-hold PET/CT. Transverse (left) and coronal (right) images of large liver metastasis in hybrid PET/CT with breath-hold CT. Center of lesion is marked with red cross on CT and with blue cross on PET. Positioning differences of liver between PET and CT acquisition resulted in mismatch of 13 mm, measured as a 3D vector.

Attenuation-Correction Artifacts

For EB_all, the average size of attenuation artifacts at the diaphragmatic liver dome in the craniocaudal direction was 11 mm (range, 0–41 mm; in 50% of cases >10 mm) and was congruent with local registration errors (Fig. 4). Visual inspection again revealed that the largest artifacts were due to expiration that was not deep enough during breath-hold CT. The occurrence of clinically relevant attenuation-correction artifacts at the diaphragmatic dome did not depend on whether the patient successfully completed the CT breath-hold instructions (error >10 mm in 45% of EB_adequate and in 56% of EB_failed; not significant). This illustrates that the artifacts are unavoidable, even when using a fast CT scanner. An example of liver deformation due to attenuation correction in breath-hold hybrid PET/CT is shown in Figure 5.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 4. 

Errors at diaphragmatic dome. Extent of image registration errors and attenuation-correction artifacts of >10 mm was comparable at location of diaphragmatic dome of liver for EB (hybrid PET/CT with breath-hold CT), FB (hybrid PET/CT with free-breathing CT), and SF (software fusion of dedicated PET and CT).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 5. 

PET/CT attenuation-correction artifacts. Hybrid PET/CT of large liver metastasis with central necrosis, with CT acquired during expiration breath-hold: coronal slices of CT (A), uncorrected PET (B), attenuation-corrected PET (C), and fused corrected PET with CT (D). Despite breathing instructions, comparison of A and B reveals a difference in diaphragm position between CT and PET acquisition. C and D demonstrate change in shape of liver on PET after attenuation correction, to falsely match CT. Liver metastasis appears partially in lung on corrected images and results from severe loss of signal intensity in region of mismatch.

For FB, the average attenuation artifact at the diaphragmatic liver dome measured 6 mm (range, 0–11 mm; 20% >10 mm), both in the cranial and caudal direction. For SF, attenuation-correction artifacts of the liver are theoretically not an issue. Control measurements at the diaphragmatic dome showed an average absolute error of 3 mm (range, 0–8 mm; thus, in all patients within 10 mm).

For all EB, FB, and SF cases, no significant attenuation-correction artifacts occurred at the lateral border or the caudal tip.

The extent of attenuation-correction artifacts at the diaphragmatic dome was significantly worse in EB_all than in FB or SF (P < 0.05). When EB_all was limited to EB_adequate, the difference with FB in the diaphragmatic dome was not significant. FB was significantly worse than SF (P < 0.05). No clinically relevant attenuation-correction errors occurred at the right lateral border and the caudal tip with either technique.

Visible Errors on acPET

At the locations of the diaphragm and the right lateral liver border, no cases showed a visually discernible mismatch of >10 mm at the liver border between acPET and CT—for all EB, FB, and SF images—regardless of the presence of registration or attenuation-correction artifacts of >10 mm. Visually discernible errors were seen at the location of the caudal tip in all image series, with values similar to the local image registration error.

Breathing Artifacts on CT

In EB_all, locoregional breathing artifacts in the liver were detected on the CT images of 4 patients (20%), all attributable to breathing motion during acquisition despite instructions to hold the breath. Furthermore, the caudal tip of the right liver lobe appeared displaced or deformed in 7 additional patients (35%), all attributable to the breathing instructions given when approaching the region of the liver in the caudocranial scanning direction. In FB, free-breathing artifacts were discernible throughout the images for all patients. No breathing artifacts were detected in the CT images used for SF. Examples of breathing artifacts on CT are shown in Figure 6. The clinically relevant (>10 mm) breathing artifacts on CT are summarized in Table 2.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 6. 

Breathing artifacts on CT: CT slices from different patients, acquired on hybrid PET/CT scanner. (A) Coronal slice of CT acquired with expiration breath-hold command. Arrows indicate artifact in middle of liver and spleen due to unsustained breath-hold. (B) Sagittal slice of CT acquired with expiration breath-hold command. Breathing commands were given relatively late and can be recognized by movement of abdominal wall (left arrow); resulting liver motion causes caudal tip of liver to appear twice (right arrow). (C) Coronal slice of CT acquired during free breathing. Breathing artifacts (arrows) are visible throughout image.

Breathing Artifacts on CT

Maintaining unforced expiration breath-hold is easily underestimated. The procedure is demanding and needs to be rehearsed before scanning. Even so, some patients will fail to sustain a breath-hold during actual scanning, causing CT artifacts (in 20% of cases in our series). Completion of a CT scan during breath-hold cannot be guaranteed, especially when scanning elderly or diseased patients. This problem is obviously related to the acquisition time for CT during whole-body scanning, as was illustrated by the absence of artifacts on the CT images that were acquired with a fast dedicated CT scanner (for SF). Another issue is the timing of breathing instructions: 33% of cases showed deformation of the caudal part of the liver, related to the deep inspiration command when approaching the midabdomen. An earlier breath-hold command would result in an increased risk on breathing artifacts toward the end of the scan, in the upper lung fields. Increasing the speed of the CT acquisition would imply thicker slices for our CT scanner, thus further degrading the quality of the low-dose CT images. This issue can be avoided by performing a separate acquisition of the liver range, as exemplified by the lack of artifacts due to the breathing protocol in SF.

As expected, the alternative strategy of free superficial breathing resulted in slight CT artifacts throughout the liver (11). In the lungs, this effect caused small lung nodules to be missed in up to 34% of cases (18). It is unknown how this translates to imaging of the liver, but such a level of missed diagnosis will not be acceptable for correlative imaging. A faster CT scanner will result in free-breathing artifacts with a lower frequency in the images, but these artifacts cannot be eliminated entirely. Thus, with regard to CT artifacts, both breath-hold and free-breathing techniques have disadvantages, and both will benefit from a faster CT scanner.

Image Registration

During the breath-hold, the exact position of the diaphragm cannot be instructed or predicted, not even in an ideal situation (i.e., with a fast CT scanner, accurate breath-hold instructions, and an exemplary patient). Furthermore, the shape of the diaphragm may differ from free breathing as during PET acquisition, because a breath-hold generates different muscle tension. This implies that differences in position and shape of the liver between PET and breath-hold CT may be unavoidable. Our results confirm that registration errors of the liver are not uncommon and are rather unpredictable in extent. Misregistrations occurred primarily in the craniocaudal direction and, in most cases, were explained by insufficient expiration during the breath-hold. Deeper expiration could not be applied, because it would increase the risk of nonsustained breath-hold. Even in cases where the compliance to breath-hold instructions was perfect, registration errors of >10 mm could occur, although the average extent of misregistration appeared lower. This may raise questions about our implementation of breath-hold instructions. Breath-hold protocols have been evaluated previously by Goerres et al. (19). They concluded that the best breathing protocol is unforced expiration breath-hold, as performed in our study, and confirmed that the impact on image registration can still be severe (relative registration errors of –25 to +19 mm vs. –16 to +44 mm in our series). Brechtel et al. have reported better values for image registration at the diaphragm, but these data seem biased because evaluation was limited to acPET images only (20).

Free breathing during CT resulted in registration errors comparable to the BH protocol (>10 mm in 30% and 40% of cases, respectively), but the maximum extent of misregistration was much lower (16 and 44 mm, respectively). The extent of misregistration was congruent with normal liver motion during free breathing as shown by Brandner et al. (21). Nakamoto et al. (22) observed even slightly worse results at the location of the right diaphragmatic dome, with 38% misregistration of >10 mm and even 10% of cases with misregistration of >20 mm (30% and 0%, respectively, in our study), with artifacts that may influence the position, shape, and apparent size of the liver on PET. Osman et al. demonstrated moderate-to-severe attenuation-correction artifacts in 18% of cases at the right diaphragmatic dome, although quantitative analysis was not performed (23). They also observed that correlation errors of liver lesions may occur, incidentally even with erroneous localization in lung instead of liver, albeit in a limited number of cases (24). Papathanassiou et al. have confirmed that lesions may be missed in liver parts that were affected by CT-based attenuation correction (25). The extent of misregistration is theoretically not dependent on the speed of the CT scanner because it “catches” the diaphragm at a random point in the breathing cycle, but this could not be evaluated in this study.

Software fusion resulted in significantly lower registration errors of both the liver, as a whole, and liver lesions. This raises the question whether such optimization of image registration and uncompromised PET image quality are possible in hybrid PET/CT. When organ-focused image registration is performed with uPET and CT images, followed by the attenuation-correction procedure, the result should be similar to SF. However, current hybrid PET/CT scanners do not provide such options. Software fusion is not ideal for high-throughput imaging, but remains useful when hybrid PET/CT is not available or for specific situations.

PET Attenuation-Correction Artifacts

Given the risk on registration errors, hybrid PET/CT with CT-based attenuation correction introduces an additional risk for artifacts on PET images. Attenuation correction will be applied erroneously on PET at the location of dense objects (26,27) or where the position of a transition from low to high photon-attenuating tissue does not correspond on PET and CT (13,22). The diaphragmatic area is very susceptible to such errors due to the sharp tissue/air transition, combined with the risk for positional differences. This may result in an apparent contour change of the liver on acPET images and reduced sensitivity for lesions in the affected area. On top of that, the PET signal will no longer be quantitative in the regions of attenuation-correction artifacts, which may compromise follow-up measurements.

In our series of EB imaging, deformation of the liver on acPET images was not uncommon and was rather unpredictable. The potential clinical impact is underlined by the presence of artifacts of >10 mm in 50% of cases. In FB, attenuation-correction artifacts were significantly less extensive but still occurred in 20% of cases with >10 mm. These problems must be considered unavoidable as long as registration errors occur and attenuation correction is performed with CT images. Dedicated PET does not readily have attenuation-correction artifacts, because the attenuation profile is measured using photons with energy identical to that of emission scanning, acquired during the same breathing pattern. No significant artifacts were detected in our series of dedicated PET images.

This leaves room for improvement of hybrid PET/CT image quality by reintroduction of 511-keV transmission imaging, although this will have no impact on image registration. Development of faster and higher-quality transmission scanning—for example, simultaneous with emission scanning (28)—is eagerly awaited and is likely to prove at least as beneficial in hybrid PET/CT as the implementation of faster CT scanners.

Selection of a Breathing Protocol

Overall, free-breathing hybrid PET/CT performs somewhat better at image registration and artifacts but has poorer image correlation due to free-breathing artifacts and missing lesions on CT. Nevertheless, both approaches have been found suitable for diagnostic correlative imaging (29). When considering a breath-hold or a free-breathing hybrid PET/CT protocol, it is important to realize that registration errors and attenuation-correction artifacts in breath-hold PET/CT can be recognized and circumvented afterward by adequate evaluation of uPET images, whereas missing small lesions on free-breathing CT is definitive. The final choice of technique may be guided by specific clinical questions, available equipment, individual patient characteristics, and personal preferences.

Alternative CT acquisition protocols all have disadvantages. Slow-CT, averaged cine-CT, and averaged multiple-series CT (as applied in external beam radiation therapy planning) may all severely degrade the image quality. Gated CT acquisition may provide excellent correlative imaging, at least when PET is acquired in gating identical to that of CT (30). Further experiments with such techniques need to be conducted.

Other PET radiopharmaceuticals for imaging of malignancy in the liver are likely to become available in the coming years. The presence of registration errors and attenuation-correction artifacts in PET/CT is independent of the PET tracer, but misregistrations will be more difficult to detect for tracers with low or no uptake in normal liver tissue. This illustrates the importance of optimized imaging and reviewing, especially for imaging of novel tracers in the near future.

The combination of PET with MRI may be preferable over PET/CT—for better soft-tissue imaging characteristics and fewer radiation dose issues—but breathing issues will remain an issue. Current MRI techniques do not allow whole-body imaging during breath-hold, and free breathing during MRI can severely distort the images. The best approach to hybrid PET/MRI is still unclear.