Abstract
For attenuation correction (AC) in PET/MRI systems, segmentation-based methods are most often used. However, the standardized uptake value (SUV) of lesions in the bone and liver, which have higher attenuation coefficients than other organs, can be underestimated, potentially leading to misinterpretation of clinical cases. Errors in SUV estimation are also dependent on the segmentation schemes used in the segmentation-based AC. In this study, this potential bias in SUV estimation using 4 different segmentation-based AC methods was evaluated for the PET/CT data of cancer patients with bone and liver lesions. Methods: Forty patients who had spine or liver lesions and underwent 18F-FDG PET/CT participated (18 women and 22 men; 20 spine lesions and 20 liver lesions; mean age (±SD), 60.5 ± 11.4 y; mean body weight, 57.7 ± 10.4 kg). The patient body region was extracted from the CT image and categorized into 5 tissue groups (air, lungs, fat, water, and bone) using Hounsfield unit thresholds, which were determined from the CT histogram. Four segmentation-based AC methods (SLA [soft-tissue/lung/air], WFLA [water/fat/lung/air], SLAB [soft-tissue/lung/air/bone], and WFLAB [water/fat/lung/air/bone]) were compared with CT-based AC. The mean attenuation coefficient for each group was calculated from 40 CT images and assigned to the attenuation maps. PET sinograms were reconstructed using segmentation- and CT-based AC maps, and mean SUV in the lesions was compared. Results: Mean attenuation coefficients for air, lungs, fat, water, and bone were 0.0058, 0.0349, 0.0895, 0.0987, and 0.1178 cm−1, respectively. In the spine lesions, the SUVs were underestimated by 16.4% ± 8.5% (SLA AC) and 14.7% ± 7.5% (WFLA AC) but not to a statistically significant extent for SLAB and WFLAB AC relative to CT AC. In the liver lesions, the SUVs were underestimated by 11.1% ± 2.6%, 8.1% ± 3.0%, 6.8% ± 3.8%, and 4.1% ± 3.8% with SLA, SLAB, WFLA, and WFLAB AC, respectively. Conclusion: Without bone segmentation, the SUVs of spine lesions were considerably underestimated; however, the bias was acceptable with bone segmentation. In liver lesions, the segmentation-based AC methods yielded a negative bias in SUV; however, inclusion of the bone and fat segments reduced the SUV bias. The results of this study will be useful for understanding organ-dependent bias in SUV between PET/CT and PET/MRI.
Hybrid imaging devices such as PET/CT and SPECT/CT are now widely used in clinical and preclinical studies. Morphologic information provided by CT is particularly useful for the localization of abnormal uptake of radiotracers and for γ-ray attenuation correction (AC) (1–3). Inaccurate AC of PET and SPECT images can lead to incorrect quantification and misinterpretation of lesions (4–9).
MRI is another important morphologic imaging tool that can be combined with PET or SPECT. MRI has several advantages over CT, including better soft-tissue contrast, no additional radiation hazard, and true 4-dimensional simultaneous imaging (9–11). It is likely that combined PET/MRI and SPECT/MRI systems will facilitate the wider use of radiotracer technologies for clinical and investigational purposes. Several research groups have been successful in demonstrating the feasibility of simultaneous PET/MRI technologies (12–15), and commercial PET/MRI systems with a gantry size sufficient for human whole-body imaging studies have recently become available (16,17).
The intensity of CT images is determined mainly by electron density, which is directly related to the photon linear attenuation coefficient. Thus, the Hounsfield units of CT can be converted into the attenuation coefficient of 511-keV γ-rays by means of a simple bilinear transformation (1–3). In contrast, MR intensity is associated with proton density or relaxation properties and is therefore not proportional to the photon attenuation power. For this reason, it is difficult to derive the γ-ray attenuation power directly from MR intensity information (18).
For PET AC using MR images, several approaches have been suggested; these methods can be categorized into template-guided AC (19,20) and segmentation-based AC (6–8). For the routine use of template-guided AC, an accurate and robust spatial normalization algorithm is required. For human whole-body PET data, a segmentation-based AC method is regarded as more reliable than template-guided AC because sometimes the latter cannot accommodate the interpatient variability of anatomic structures in the human torso (6,7,9,18,21).
For segmentation-based AC methods, we segment and categorize tissue groups on the basis of MRI intensity. Usually, 3 (soft-tissue, lungs, and air; no distinction between water-equivalent and fat tissues (6)) or 4 (water-equivalent, fat, lungs, and air; water and fat distinction using Dixon MRI sequence (7)) segmentation-based AC methods are used for torso images. However, bone segmentation has not been as successful in MRI-based torso PET AC until now because it is hard to distinguish bone and air intensities in MR images acquired using standard pulse sequences (9). The relaxation time of cortical bone structures is too short for signal measurements and is a major cause for the similar image intensities of cortical bone and air. Consequently, bone structures were included within the soft-tissue regions in the previously discussed 3 or 4 segmentation-based AC methods.
Recently, several groups have investigated the segmentation of bone structures from soft tissue using the ultra-short time echo (UTE) sequence (22–24). Using the UTE sequence, we can measure early relaxation signals from cortical bone structures. However, more than a 20% error was introduced in the reconstructed PET images by AC using MR images acquired with the UTE sequence, and some boundary regions between soft tissue and air in the segmented MR images were misinterpreted as bone structures (23,24). The UTE sequence was useful to segment bone only for the head, a relatively simple structure compared with the torso (22).
Using segmentation-based AC methods without bone segmentation, we can expect quantitative inaccuracies in attenuation-corrected PET images (25). In organs with relatively high photon attenuation, uptake of radiotracers would be underestimated. In our preliminary evaluation of organ-dependent bias in PET tracer uptake, we found that uptake in liver and bone was underestimated in segmentation-based AC regardless of inclusion of bone segmentation (26). Therefore, in this study, we evaluated the biases in standardized uptake value (SUV) of 18F-FDG PET with segmentation-based AC in patients who had tumor lesions in bone or liver. The accuracies of segmentation-based AC methods used in current commercial PET/MRI systems (SLA [soft-tissue/lung/air] and WFLA [water/fat/lung/air]) were compared with CT-based AC. Moreover, 2 potential MRI-based AC methods (SLAB [soft-tissue/lung/air/bone] and WFLAB [water/fat/lung/air/bone]) were included in the comparison under the assumption that the bone structures could be segmented from the soft-tissue group. Abbreviations, number of segments, and related MRI sequences are summarized in Table 1.
Abbreviations of Segmentation-Based AC Methods for PET/MRI That Were Compared in This Study
MATERIALS AND METHODS
Subjects and Data Acquisition
The whole-body PET/CT scan data of 20 liver (8 women and 12 men) and 20 spine (10 women and 10 men) oncology patients were retrospectively analyzed. Mean (±SD) patient age and weight were 60.5 ± 11.4 y (range, 28–81 y) and 57.7 ± 10.4 kg (range, 37–80 kg), respectively. The retrospective use of the scan data was approved by the Institutional Review Board of Seoul National University Hospital, Seoul, Korea.
A Biograph TruePoint TrueV PET/CT scanner (Siemens Medical) was used in this study to acquire PET and CT datasets. The scanner comprised 4 rings of detector blocks. Each ring contained 48 detector blocks, and each detector block consisted of a 13 × 13 array of lutetium oxyorthosilicate scintillation crystals with an individual dimension of 4 × 4 × 20 mm. The axial field of view of the PET scanner was 216 mm. The scanner was operated in 3-dimensional mode for PET emission scans. In all patients, the upper body from neck to upper thigh was covered by a 5- or 6-bed-position emission scan. The duration of the emission scan was 2 min/bed position after 60 min after injection of 18F-FDG (301 ± 55 MBq, 5.18 MBq/kg).
Segmented AC Map
To investigate the SUV bias due to segmentation-based AC, the segmentation-based attenuation maps were generated from the CT images. First, the portion of the CT image representing the patient body was extracted. The Hounsfield unit threshold (−400) was applied to the CT image to extract the patient body and bed automatically, and the patient bed was manually separated from the patient body. Then, the Hounsfield unit histograms were obtained from each patient’s CT images, and thresholds between each segment (air, lungs, fat, water, and bone) were determined at the local minima of the histogram. Using the thresholds, patient body parts were automatically classified into 3, 4, or 5 segments (SLA, WFLA, SLAB, and WFLAB). The mean Hounsfield units for each segment were then calculated from 40 CT images and assigned to the body segment of each patient. Finally, the patient bed was added. These images are compared in Supplemental Figure 1 (supplemental materials are available online at http://jnm.snmjournals.org).
Image Reconstruction
All emission data were reconstructed using these 5 different attenuation maps (CT, SLA, WFLA, SLAB, and WFLAB) and 2-dimensional ordered-subsets expectation maximization with 2 iterations and 14 subsets after random, scatter, attenuation, and normalization corrections and data rebinning. PET reconstruction software (e7 tools) from Siemens Healthcare was used for this purpose. Reconstructed images had dimensions of 168 × 168 × 405 for the 5-bed-position scan and 168 × 168 × 479 for the 6-bed-position scan, with 4.073-mm transaxial pixel spacing and a 2.027-mm axial slice interval.
SUV Comparison
A region of interest was drawn on metastatic or primary tumor lesions for each PET dataset. For each patient, a single region of interest was applied on a tumor lesion containing the highest-SUV voxel and covering the entire lesion volume. The same region of interest was used for all attenuation-corrected PET images, and mean SUVs were obtained. The SUVs from the segmentation-based AC methods (SLA, SLAB, WFLA, and WFLAB AC) were compared with CT AC by evaluating their percentage differences.
Statistical Analysis
Paired t tests were performed on SUVs between CT AC and each segmentation-based AC method. Bonferroni adjustment was applied to compensate for multiple comparisons. A P value of 0.05 was used to determine statistical significance.
RESULTS
For all tissue groups, the mean attenuation coefficients for 511-keV γ-rays were calculated from the data of 40 patients. The mean attenuation coefficients of air, lung, soft-tissue (fat and water), and bone segments were 0.0058, 0.0349, 0.0952 (0.0895 and 0.0987), and 0.1178 cm−1, respectively.
Figure 1 shows difference images of SUV between segmentation-based AC and CT AC. Without bone segmentation, the spine region was considerably underestimated (Figs. 1A and 1C). Another remarkable finding was that the liver was also underestimated, especially in segmentation-based AC without a distinction between water and fat (Figs. 1A and 1B).
SUV difference images between SLA and CT AC (A), between SLAB and CT AC (B), between WFLA and CT AC (C), and between WFLAB and CT AC (D).
Examples of attenuation-corrected PET images are shown in Figures 2 and 3. Figure 2 shows images of a patient with multiple bone metastases for which attenuation was corrected by the 5 different AC maps. Without bone segmentation, 18F-FDG uptake in the lesions on the spine was significantly underestimated (Figs. 2B and 2D); however, the differences do not seem to be significant with bone segmentation (Figs. 2C and 2E). Figure 3 shows images of a patient who had a hypermetabolic lesion in the right lobe of the liver for which attenuation was corrected by the 5 different AC maps. Segmentation-based AC underestimated 18F-FDG uptake in comparison with CT-based AC. The error in SUV was greater in segmentation-based AC methods without a water–fat distinction (Figs. 3B and 3C).
Attenuation-corrected PET data of patient with multiple bone lesions using CT (A), SLA (B), SLAB (C), WFLA (D), and WFLAB (E). Values below marked lesions are percentage differences in SUV relative to CT AC.
Attenuation-corrected PET data of patient with hepatic lesion using CT (A), SLA (B), SLAB (C), WFLA (D), and WFLAB (E). Values below marked lesions are percentage differences in SUV relative to CT AC.
For all patient data, the percentage differences in SUV (SLA vs. CT, SLAB vs. CT, WFLA vs. CT, and WFLAB vs. CT) were quantified (Fig. 4). In spine (bone) lesions, the SUVs were underestimated by 16.4% ± 8.5% (SLA, P < 10−6) and 14.7% ± 7.5% (WFLA, P < 10−7) with segmentation-based AC without bone segmentation, but the difference was not statistically significant with segmentation-based AC with bone segmentation (P > 0.05 for both SLAB and WFLAB). In liver lesions, the SUVs were underestimated by 11.1% ± 2.6% (P < 10−9), 8.1% ± 3.0% (P < 10−8), 6.8% ± 3.8% (P < 10−6), and 4.1% ± 3.8% (P < 10−3) with SLA, SLAB, WFLA, and WFLAB AC, respectively.
SUV percentage differences in segmentation-based AC compared with CT AC for bone and liver lesions.
DISCUSSION
AC is a necessary process for the reconstruction of PET images. In particular, accurate AC is essential for quantitative studies such as of tracer kinetics or radiation dosimetry and for clinical follow-up studies in which the severity of cancer is evaluated using the mean or maximum SUV. However, in PET/MRI systems, the AC map cannot be obtained directly from MRI because voxel intensities in MR images are not related to photon attenuation coefficients and are not quantitative (i.e., they are expressed in arbitrary units) (21). For PET/MRI combined systems, therefore, attenuation coefficients have to be estimated and not measured. The most reliable candidate method for MRI-based AC is segmentation-based AC. This method assumes homogeneous attenuation power in a segment. However, the segmentation-based AC method distinguishes only a few regions, and as a result, each segment contains multiple organs of differing photon attenuation. Organ-dependent bias is the result of this assumption.
For segmentation-based AC without bone segmentation, we observed considerable negative bias (−16.4% and −14.7% for SLA AC and WFLA AC, respectively) in spine cancer lesions mainly because the bone structures, which have high-attenuating components, were regarded as soft tissues, which have low-attenuating components. In a previous investigation on the effects of bone segmentation using beagles (8), 3-segment–based AC yielded approximately 10% bias in the spine region, and this bias decreased to 3% if bone segmentation was applied. In addition, there was a recent report that substitution of bone by soft tissue in the data of 22 patients introduced tracer uptake underestimations in osseous regions by 11.2% (27). In this study, we confirmed this bone segmentation effect in human data. Therefore, bone segmentation is a fundamental process for accurate quantification of bone regions.
The UTE sequence provides the possibility of accurate bone segmentation. Until recently, however, the UTE sequence was not a perfect solution for bone segmentation, especially in the torso. Images obtained with UTE have high noise because of the short relaxation time of cortical bone structures. To improve image quality, the acquisition time should be lengthened because repetitive acquisitions are required. Another main limitation for UTE in body imaging is the difficulty in achieving a homogeneous magnetic field across a large field of view. Moreover, streak artifacts and misinterpretation of body edges as bone are further challenges in using UTE sequences for AC of PET images (22,24). All these limitations are more severe for thorax and pelvis imaging than for head and neck imaging. Consequently, most UTE studies are focused on brain studies (22–24).
In this study, we also evaluated SUV in the liver. The results shown in Figures 3 and 4 indicate that both bone segmentation and water–fat distinction contribute to the reduction of SUV bias in the liver. Although the combination of these 2 techniques (WFLAB AC) reduced bias to less than 5%, the negative bias was statistically significant. This bias might be caused by the same processes as in bone lesions corrected with segmentation-based AC without bone segmentation. The density of the liver is relatively higher than that of other tissue organs. The mean attenuation coefficient of the liver calculated from a region of interest on the liver in our 40 subjects was higher (0.1018 cm−1) than that of the water segment. Thus, organ-dependent bias according to organ density will exist if various soft tissues are regarded as a single tissue group. Organ-dependent bias would be an unavoidable limitation in quantification studies using segmentation-based AC in PET/MRI.
The assignment of a nonzero attenuation coefficient (0.0058 cm−1) to air in this study would not reflect reality and reduce the bias with respect to the reference CT reconstruction.
CONCLUSION
Segmentation-based AC without bone segmentation caused a considerable underestimation of the SUV of spine lesions; however, with bone segmentation, the bias was acceptable. In liver lesions, segmentation-based AC methods yielded negative bias in SUV; however, inclusion of the bone and fat segments reduced the SUV bias. The results of this study will be useful for understanding the organ-dependent bias in SUV between PET/CT and PET/MRI.
DISCLOSURE STATEMENT
The costs of publication of this article were defrayed in part by the payment of page charges. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.
Acknowledgments
This work was supported by grants 800-20110287, 2008-2003852, 2010-0026012, and R31-10089 (WCU program) from the Ministry of Knowledge Economy and KOSEF, funded by the Ministry of Education, Science, and Technology. No other potential conflict of interest relevant to this article was reported.
Footnotes
Published online Oct. 18, 2012.
- © 2012 by the Society of Nuclear Medicine and Molecular Imaging, Inc.
REFERENCES
- Received for publication February 6, 2012.
- Accepted for publication July 24, 2012.
Jump to section
Related Articles
Cited By...
- Generation of PET Attenuation Map for Whole-Body Time-of-Flight 18F-FDG PET/MRI Using a Deep Neural Network Trained with Simultaneously Reconstructed Activity and Attenuation Maps
- Improving the Accuracy of Simultaneously Reconstructed Activity and Attenuation Maps Using Deep Learning
- MRI-Based Attenuation Correction for PET/MRI Using Multiphase Level-Set Method
- Segmentation-Based MR Attenuation Correction Including Bones Also Affects Quantitation in Brain Studies: An Initial Result of 18F-FP-CIT PET/MR for Patients with Parkinsonism