Hiroshi Kawaguchi1,2, Tkayuki Obata2,3, Hiromi Sano2, Eiji Yoshida2, Mikio Suga4, Yoko Ikoma2, Yukari Tanikawa1, and Taiga Yamaya2
1Human Informatics Research Institute, National Institute of Advanced Industrial Science and Technology (AIST), Tsukuba, Japan, 2Molecular Imaging Center, National Institute of Radiological Sciences, Chiba, Japan, 3Research Center for Charged Ion Therapy, National Institute of Radiological Sciences, Chiba, Japan, 4Center for Frontier Medical Engineering, Chiba University, Chiba, Japan
Synopsis
The current
attenuation correction method for human pelvic PET/MRI contains several
problems such that the attenuation due to bone is not considered and a specific
MR imaging, intended for attenuation correction only, is needed. In this study,
we proposed a method to generate the distribution of attenuation correction
factors with considering the bone attenuation using diagnostic T1-weighted MRI
for pelvic PET/MRI scanning. The proposed method is the hybrid of the tissue
segmentation based on Gaussian mixture model and the non-liner registration of
tissue probability to a subject image. The simulation results showed that
attenuation correction using the proposed hybrid method reduced the error on
PET image than the conventional method.Introduction
The ability to make quantitative
measurements of radioactivity is essential for human PET. Although several
MR-based attenuation correction (AC) methods have been reported to estimate the
spatial distribution of the attenuation coefficient (μ-map) for PET/MRI
scanners, a few methods were applicable for the body trunk
1. Especially,
the conventional AC for
pelvic PET/MRI contains several problems such that the attenuation due to bone
is not considered and a specific MR imaging, intended for AC only, is needed. In this
study, we proposed a method to generate theμ-map with considering the bone
attenuation using the diagnostic T1-weighted MRI for the pelvic PET/MRI scan.
Methods
Fig. 1 is a schematic of PET/MRI AC with the proposed hybrid-segmentation atlas method (HSAM). The algorithm is
based on two probability maps: the intensity-based probability map, P, constructed
by the soft segmentation of subject’s T1-weighted image (T1WI) and the tissue
probability atlas, Q. As shown in Fig .2, the pixel intensity roughly divided
into 3 groups in the pelvic T1WI. The first step in making P is to determine
the threshold value to distinguish very low intensity pixels from the others. The
smallest local minimum on the intensity histogram was used as the threshold
value. Next, the histogram of the remained voxels was fitted to three-component
Gaussian mixture model and intensity-based probability maps were produced. The
first step in making Q is to construct the template of the T1WI and the tissue
probability map from the database of images. Next, the affine transformation was
computed to register the T1WI template to individual T1WI. In addition, the
demons method was applied to improve the accuracy of the registration
2.
Next, these transformations were applied to each Q. The complete tissue probability
maps, Pair, Pbone and Psoft, were culculated
from P and Q’ using equations shown in Fig. 1. Finally, the most suitable µ-value
was assigned to each voxel following complete tissue probability maps.
The MR-based AC methods were evaluated using MRI
scans from 20 prostate cancer patients. All 3D T1WI were acquired with the Achieva 1.5T
scanner (Philips Healthcare, Best, Netherlands) using the multi-slice spin echo
sequence (orientation: transaxial, TR: 675 ms, TE: 10 ms, matrix size: 512x512,
FOV: 350 mm, NEX: 2, slice thickness: 4 mm, no gap). The proposed method was
validated by leave-one-out method, i.e. templates were constructed from 19
subjects except the subject for evaluation. In addition to HSAM, µ-maps were also
generated by the conventional segmentation-based method (SBM) and atlas-based
method (ABM) as described in Fig. 1. The digital phantom to simulate the PET
scan was constructed by the manual segmentation of the subject’s T1WI. Jaccard
index, JI=(A^B)/(AvB), was calculated between the µ-map of digital
phantom (A) and each MR-based µ-map (B). The PET emission images were
reconstructed with two-dimensional filtered back-projection after AC. The error (E) in
the PET image was calculated with the following equation: E = ∑
N|I – I
ref|/N, where I is the pixel
intensity in the ROI on the reconstructed image being evaluated, I
ref is that on the reference
image and N is number of pixels in the ROI. The reference image was
taken to be the PET image corrected by µ-map of the digital phantom. The statistical
difference between MR-based AC methods was evaluated by the one-way repeated measure ANOVA with Bonferroni
correction (α=0.05).
Results and
Discussion
Figure 3
shows the MR-based µ-maps and reconstructed PET images of a typical subject. Jaccard
indices and error in PET image were summarized in Table 1. While SBM and ABM
could not distinguish the bone and the intestinal gas regions in the µ-map,
respectively, HSAM succeeded to segment most of these regions. Jaccard indices for
the HSAM were statistically better than that for ABM in all regions and SBM in
the soft tissue and the bone. In the PET image, HSAM and ABM succeeded to
reconstruct the radioactivity of the prostate region, while the SBM produced underestimation
due to the bone attenuation. The false positives are found on the boundary of
the two different regions, which will be caused by the incorrect segmentation
and the misregistration of the template.
Conclusion
The fully
automated µ-map generation method was proposed for the pelvic PET/MRI scan
using the diagnostic T1WI. The proposed HSAM provided more accurate
segmentation results than conventional SBM and ABM. The AC using the µ-map from
HSAM and ABM provided more accurate PET image than that by SBM, which indicates
bone attenuation should be considered for the AC of the pelvic PET.
Acknowledgements
This work was partially supported by the KAKENHI grant for HK(15K15216)References
1. Keereman
V, Mollet P, Berker Y, et al. Challenges and current methods for attenuation
correction in PET/MR. MAGMA, 2013:26(1):81-98.
2. Vercauteren T, Pennec X, Perchant A, et al.
Diffeomorphic demons: efficient non-parametric image registration. Neuroimage.
2009:45(1 Suppl):S61-72.