An atlas-based automatic brain segmentation method for regional SUV quantification in hybrid PET/MR

Background ： Quantitative analysis of brain positron-emission tomography (PET) depends on structural segmentation, which can be time-consuming and operator-dependent when performed manually. Previous automatic segmentation methods usually fitted subjects’ images onto an atlas template for group analysis, which changed the individuals’ images and affected regional PET segmentation. We proposed an automatic segmentation method, registering atlas template to subjects’ images (RATSI), which created an individual atlas template and may be more accurate for PET segmentation. We segmented two representative brain areas in twenty Parkinson disease (PD) and eight multiple system atrophy (MSA) patients performed in hybrid positron-emission tomography/magnetic resonance imaging (PET/MR). The segmentation accuracy was evaluated using the Dice coefficient (DC) and Hausdorff distance (HD). and the standardized uptake value (SUV) measurements of these two automatic segmentation methods were compared, using manual segmentation as a reference. Conclusion ： The RATSI was more accurate for the caudate nucleus and putamen automatic segmentation, and can be used for regional PET analysis in hybrid PET/MR.


Background:
Positron emission tomography (PET) is a molecular imaging method that uses the annihilation reactions of different positron emitters in radiotracers to generate images from the 511 keV gamma rays emitted. The positron emitters used, such as 18 F and 11 C, are bound to physiologically active substances to detect disease. Because physiologic images are relatively low in spatial resolution, computed tomography (CT) has been combined with PET for both anatomical location and attenuation correction. PET/CT scanning is widely applied in the evaluating of tumors, cardiac disease, CNS disorders, and infection/inflammation [1,2].
Compared with CT, magnetic resonance imaging (MRI) has better soft tissue contrast and can obtain multiparametric images, e.g., T1-weighted images (T1WI), T2-weighted images (T2WI), proton density-weighted images (PDWI), or diffusion-weighted imaging (DWI) without ionizing radiation exposure. Therefore, the combination of PET and MRI should provide much more functional and structural information than CT without CT's contribution to the overall radiation dose [3][4][5]. Combining PET with MRI was initially considered in the 1990s [6,7]. Until 2010, with the advent of a magnet-compatible avalanche photodiode detector (APD), the first commercial whole-body hybrid PET/MR system (Siemens Biograph MRI scanner, Siemens Healthcare, Erlangen, Germany) [8] was introduced. In 2014, the latest generation of clinical whole-body hybrid PET/MR scanner (SIGNA PET/MR, GE Healthcare, Waukesha WI, USA) emerged with silicon photomultipliers (SiPMs), which digitize and process the signal directly within the magnetic field, which resulted in a thousand-fold improvement in time resolution, permitting time-of-flight (TOF) imaging [9]. 5 PET/MR can simultaneously obtain images from the two modalities and is very useful for applications in neuroimaging. Catana et al. summarized the potential clinical application of PET/MR in patients with neurological disorders [10]. Previous studies have shown that the metabolic patterns of Parkinson's disease (PD) and atypical Parkinson's syndrome, e.g., multiple system atrophy (MSA), progressive supranuclear palsy (PSP), and corticobasal degeneration (CBD), were different [11]. However, it is usually difficult to quantitatively analyze abnormalities in different brain regions accurately [12]. Because of PET's relatively low spatial resolution and the complexities of brain anatomy, the analysis of regional PET quantification relies on MRI. Brain structural segmentation based on MRI is very useful to localize regional CNS metabolism for clinical diagnosis.
Segmentation is one of the fundamental challenges in biomedical image analysis in that brain morphological characteristics are very complicated. This has been widely investigated to help diagnosis or surgery, e.g., deep brain stimulation (DBS) [13,14]. Brain structural segmentation can be performed manually or automatically. Manual segmentation is highly time-consuming, requires expert anatomical knowledge [15,16], and is subject to operator-dependence, especially when the signal-to-noise ratio or resolution is suboptimal.
However, manual segmentation is still often employed and the results are always used to establish a valid ground truth against which to assess automated segmentation results [17].
Automatic segmentation methods depending on algorithms can be very convenient, and the results are usually objective and reproducible [17]. Automatic segmentation methods must 6 show accurate imaging co-registration, including the co-registration of images from different modalities, and accurate co-registration with a common reference template, such as the Montreal Neurological Institute (MNI) brain template [18].
For PET quantitative analysis, previous studies usually extracted the regional cerebral standardized uptake values (SUVs) directly based on an atlas, e.g., the widely used 3D stereotactic surface projection (3D SSP), or acquired the PET/CT and MRI data sequentially, registered the PET and MRI by postprocessing [19,20], and performed quantitative SUV analysis. In hybrid PET/MR, the two modalities are acquired simultaneously, which avoids image misregistration. The main difficulty is registering the subjects' images with a referenced brain atlas. Many studies have focused on developing or optimizing algorithm performance for cortical structures in support of the functional MRI literature [21,22].
However, the evaluation of effects of different automatic segmentation methods on PET SUV quantification is lacking. To label brain regions for group analysis, automatic segmentation (herein referred to as "the traditional method") always fits the subjects' images to a common reference space [23,24], and PET quantification is performed in the transformed space, which fits the individual image data to adapt to the common reference frame and probably influences the accuracy of regional SUVs [25].

Materials and methods
In this study, we inversely registered the atlas in the template space to the original PET/MR data space and calculated the SUV-defined as registering atlas template to subjects' images 7 (RATSI) method-then compared the quantification with that of the traditional method of fitting the clinical image data to the template space. Using the two automatic methods, we segmented two representative brain areas containing four regions: the left caudate (CAU_L), right caudate (CAU_R), left putamen (PUT_L), and right putamen (PUT_R) in twenty PD patients, then compared the SUVmean and SUVmax in the corresponding brain regions. The manual segmentation method was also performed and used as the ground truth. The Dice coefficient (DC) and Hausdorff distance (HD) were used to evaluate inter-rater variability.
The RATSI method was applied to quantify the differences in 18 F-FDG uptake between PD and MSA groups in multiple brain regions, including caudate, putamen, and the cerebellar gray matter.

Subjects and Data
We retrospectively studied patients who had undergone 18  Twenty typical PD (60 ± 5 y) and eight MSA patients (60 ± 8 y) were involved in this study.
The diagnosis was according to the diagnostic criteria for PD in China in 2016 [26] and the MSA diagnostic criteria of a Chinese expert consensus in 2017 [27]. The exclusion criteria were as follows: (1) a clear history of stroke, with brain MRI examination revealing large 8 cortical infarction or hemorrhagic manifestations; (2) CNS infectious disease; (3) brain tumors or history of head trauma; (4) history of craniocerebral surgery; and (5) suboptimal image quality.

Image Acquisition and Reconstruction
All patients underwent 18 FDG-PET and MRI brain imaging simultaneously in a hybrid PET/MR scanner (3.0 T, SIGNA TOF-PET/MR, GE Healthcare). The 18 F-FDG was produced in our center by a Minitrace cyclotron (GE Healthcare, USA) and automatic synthesizer (PAT Biotechnology Company, Beijing, China). The radiochemical purity was > 95%.
All participants fasted for at least 6 h and stopped any drugs that could affect brain glucose metabolism for at least 12 h before the 18  correction was atlas-based MRI attenuation correction, combined with Dixon water-fat separation methods [28].

Brain Segmentation and SUV Quantification
Automatic brain segmentation was based on an atlas template from the automated anatomical labeling atlas (http://www.gin.cnrs.fr/en/tools/aal-aal2/) shown in Figure 1. There are 70 segmented regions labeled from 1 to 70 in this brain atlas, which were used for both the traditional method and RATSI. By registration of 3D T1-weighted MRI to MNI space with SPM12 segmentation (http://www.fil.ion.ucl.ac.uk/spm/download/), an inverse deformation field could be produced. The RATSI fitted the acquired inverse deformation field to the brain atlas template, generating a personalized brain atlas for every subject, which then was used for regional 18  For the RATSI automatic brain segmentation, the detailed steps were as followed: 1. The medicine (DICOM) format of 3D T1 MRI and 18 F-FDG PET images was converted to the neuroimaging informatics technology initiative (NIfTI) format using SPM12 for following processing.
2. The 3D T1 images were normalized to MNI space. The results would produce the inverse deformation field, which extracted the information of transformation between the data acquisition space and MNI space.
3. By utilizing the inverse deformation filed, the anatomical labeling atlas in the MNI space was transformed into the data acquisition space and produced the personalized atlas template, which was in the same space with FDG images. 10 4. The regional FDG images were segmented according to the personalized atlas template and further, the mean or max SUV values in different regional brain could be calculated.

Inter-rater Reliability
The four brain regions were segmented with the manual method on twenty PD subjects for inter-rater variability evaluation by using the parameters DC and HD. The DC evaluates the similarity between two volumes by measuring their overlap [30].
where A and B represent the segmentation volumes of the automatic methods and manual method, respectively. A∩B represents the intersection of the two volumes. A DC value of 1 represents two identical segmentations while a DC value of 0 represents no overlap between the two segmentations. HD usually measures how far two subsets of a metric space are from each other, and here, determines on average how much the two segmented volumes differ. A smaller HD represents a closer agreement between two volumes.

Statistical Analysis
The differences in parameters DC and HD were analyzed by a paired t-test. The four segmented brain regions used as binary masks were overlapped on the PET images to extract the regional SUVmean and SUVmax. One-way analysis of variance (ANOVA) was used to compare the differences in quantitative SUVs among the three segmentation methods. The F-test was used to test whether the variance was homogeneous and the two-tailed t-test was used to compare the differences in SUVs in the different regions in the basal ganglia and cerebellar gray matter between the PD group and MSA group. P>0.05 was considered variance homogeneous for F-test. P < 0.05 was considered statistically significant for t-test.

Results
The DC and HD for the traditional method (green boxplots) and RATSI method (red boxplots) are displayed in Figure 3. The mean DCs of RATSI were much larger than those of the traditional method, while the mean HD of RATSI was much smaller. The quantitative values are listed in Table 1. The corresponding two-tailed t-test results are also shown. Significant differences in DC and HD were found between the two methods (P < 0.05). The maximum 12 DC was nearly 0.8 for right caudate nucleus segmentation in the RATSI method. Figure 4 shows the representative segmented caudate and putamen ROIs on the left and right side with the three segmentation methods, which were overlain on T1WI and displayed with different colors for better visualization, including coronal, sagittal and axial views. Visually, the segmented volumes from the traditional method were larger than the manual and RATSI segmentation volumes, especially for the right caudate nucleus as indicated by the white arrows.
The distributions of SUVmean and SUVmax extracted with the three segmentation methods in the caudates and putamina from all twenty PD patients are box-plotted in Figure 5. For SUVmax, the quantification was nearly the same, while the SUVmean of both the automatic segmentation methods were slightly lower than those of the manual segmentation. The corresponding quantitative SUV and ANOVA results are listed in Table 2. No significant differences were found in SUVmean or SUVmax among the three segmentation methods (P>0.05).
The consistencies of SUVmean obtained from manual and automatic methods were evaluated with Bland-Altman plots. As shown in Figure 6, the transverse and longitudinal axis represent, respectively, the mean and differential values calculated by the two automatic methods. Most (94%) of the dots were within the two 95% consistency limit lines, which indicated that it is feasible to measure the SUVmean with the automatic segmentation method based on the atlas 13 template.
The SUVmean extracted with RATSI in brain nuclei between the PD group and MSA group are summarized in Table 3. All the SUVmean were decreased in the MSA group compared with those in the PD group. The difference in SUVmean in the cerebellar gray matter was statistically significant between the two groups (P<0.05).

Discussion
The quantitative analysis of regional cerebral metabolism is very meaningful for diagnosis of nervous system diseases and exploring brain function. In this study, by using hybrid PET/MR, the registration of PET functional images and MRI structural images was avoided.
Furthermore, the personalized templates, which were finally used for 18 F-FDG PET regional segmentation and quantification, were built based on an inverse deformation field obtained by registering the atlas template to 3D T1WI in SPM12. Compared with the traditional method, the RATSI could acquire larger DCs, more consistent with those of the manual method.
Therefore, the RATSI registration method is probably advantageous for PET quantitative analysis, especially for point-to-point image analysis. In addition to 18 F-FDG PET, this method could also be used for cerebral PET analysis with other tracers.
The ability to distinguish between PD and Parkinson syndrome is very difficult, especially in the early stage. Previous reports suggested the diagnostic accuracy rate of early PD was only 53% [31]. In our study, the RATSI method was used for calculation and comparisons of 14 regional brain SUVs between the PD and MSA groups. The results suggested the method can effectively quantify the regional SUVs and find the differences of 18 F-FDG PET metabolic pattern between the PD and MSA groups. The SUVmean was significantly lower in the cerebellar gray matter in the MSA group compared with the PD group. The results are consistent with previous clinical studies and meta-analyses that use the metabolic brain network based on 18 F-FDG-PET for the differential diagnosis of Parkinson syndrome [32,33].
The quantitative evaluation of segmentation accuracy with DC and HD suggested the RATSI can generate better segmented results than the traditional method. However, the SUVmax were not significantly different, due to that the SUVmax usually was less related with the edge of segmented regions. Also, there were no significant differences in the SUVmean among the three segmentation methods in the caudate and putamen. We speculate that becasuse the caudate and putamen are relatively large and contain so many voxels, the marginal differences caused by the two automatic segmentation methods had little influence on the SUVmean. Therefore, the RATSI method probably would be more advantageous for smaller areas.
The personalized atlas template was very important for segmentation in the RATSI method, which depended on the deformation field. In this study, for generating the deformation field consistently, the default parameters were used on the SPM12. The parameters for every registration should be optimized to acquire more precise and personalized deformation field data.
In this study, we focused on manual and automatic segmentation and quantification evaluation of the caudate and putamen. The automatic method could be theoretically extended to other structures according to the atlas templates. However, some deep brain structures show poor contrast and are typically difficult to be visualized on T1WI, e.g., the red nucleus and substantia nigra, but are better visualized on T2WI. Fonov et al. (Fonov V et al. 2011) and Xiao et al. (Xiao et al. 2015) demonstrated better segmentation results from non-rigidly warping the T1WI and T2WI to a common template space. In the future, multi-modality data, that is, T1WI, T2WI, PDWI, might be used for automatic segmentation. This needs to be explored in the future.

Conclusions
We proposed an automatic segmentation method, registering atlas template to subjects' images (defined as RATSI), which could be used for regional PET analysis. Comparing with the traditional method, the RATSI was more accurate for the caudate nucleus and putamen automatic segmentation, while has little effects for their max and mean SUV calculation in hybrid PET/MR. And it could be theoretically extended to other structures according to the atlas templates.

Consent for publication
Not applicable

Availability of data and materials
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request  The brain atlas template, which segments the brain into 70 regions, labeled with numbers from 1 to 70 and showed with different colors. It was used for the following automatic segmentations.

Figure 2
The diagram displaying the processing steps of the RATSI method for whole brain automatic segmentation and regional 18F-FDG PET quantification.