A total of 51 adult patients (age > 20 years) with newly diagnosed and histologically confirmed LGGs (World Health Organization [WHO] grade II or III) who underwent FDOPA PET and MRI between 2007 and 2019 were retrospectively selected. All patients were diagnosed with LGGs according to the WHO 2007 or 2016 classification, and were classified based on IDH mutational status and 1p19q codeletion status by conventional techniques . No patients underwent stereotactic biopsy prior to FDOPA PET or MRI. OS was measured from the time of the PET scan until death or the censored date (maximum, 2000 days) with median term of 740 days. All patients signed institutional review board-approved consent forms to have their imaging, clinical, and molecular data included in our research database (IRB IRB#10-000655). The patient cohort in this study was partly overlapped with a previous study .
FDOPA PET Image Acquisition
FDOPA PET images were acquired with a high-resolution full-ring PET/CT scanner (ECAT-HR; CTI/MIMVista; Siemens, Knoxville, TN, USA) after the subjects fasted for more than four hours. Following previously reported procedures, FDOPA was synthesized and injected intravenously [22, 23]. Computed tomography images were acquired prior to the PET scan for attenuation correction. Three-dimensional FDOPA emission data were acquired for a total of 30 minutes, and the data were integrated between 10–30 minutes following the injection to obtain 20-minute static FDOPA images after reconstruction. FDOPA PET images were reconstructed using an ordered-subset expectation maximization iterative reconstruction algorithm consisting of six iterations with eight subsets [24, 25]. Finally, a Gaussian filter with a full width at half maximum of 4 mm was applied. The resulting voxel sizes were 1.34 mm × 1.34 mm × 3 mm for FDOPA PET images. Standardized uptake value (SUV) maps of FDOPA were calculated based on the radioactive activity divided by the decay-corrected injected dose per body mass . Resulting SUV maps were subsequently normalized relative to the median value of the normal-appearing striatum (nSUV) [20, 27].
Magnetic Resonance Image Acquisition
Anatomical MRI consisted of standard T1-weighted pre- and post-contrast images (2D axial turbo spin echo with 3-mm slice thickness and no interslice gap, or 3D inversion-prepared gradient echo images with 1–1.5 mm isotropic voxel size) and T2-weighted fluid-attenuated inversion recovery (FLAIR) images acquired at 3-mm slice thickness with no interslice gap using a 1.5-T or 3-T clinical MRI scanner.
Postprocessing and ROI Analysis
The processing procedures are described in Fig. 1. A single tumor region of interest (ROI) was segmented based on the regions of hyperintensity on T2-weighted FLAIR images by a board-certificated neuroradiologist (H.T. with 14 years of clinical experience) with Analysis of Functional NeuroImages software (AFNI; NIMH Scientific and Statistical Computing Core; Bethesda, MD, USA; https://afni.nimh.nih.gov) using a semi-automatic procedure as previously described [28, 29]. FLAIR and PET images, as well as FLAIR hyperintense ROIs were registered to the post-contrast T1-weighted images for each patient using a six-degree of freedom rigid transformation and a mutual information cost function using FSL software (flirt; FMRIB, Oxford, UK; http://www.fmrib.ox.ac.uk/fsl/). Each registered ROI was applied to the corresponding PET images. Maximum nSUV (nSUVmax) within the FLAIR hyperintense ROI and biological tumor volume (BTV), which included the voxels within the ROI higher than the median uptake value in the striatum, were calculated. The LGG patients were stratified into FDOPA hypometabolic (nSUVmax < 1) and hypermetabolic (nSUVmax > 1) groups according to the nSUVmax with a cut-off value of one relative to the striatum. This cut-off value, one, is determined according to the previous suggestion . Anatomical FLAIR volume and BTV are reported as milliliters.
Each post-contrast T1-weighted images was registered to a 1.0 mm isotropic T1-weighted brain atlas (MNI152; Montreal Neurological Institute [MNI]) using a 12-degree of freedom affine transformation with statistical parametric mapping 12 software (SPM12; Wellcome Trust Centre for Neuroimaging, London, UK; https://www.fil.ion.ucl.ac.uk/spm/software/spm12/), and applied the transform matrix to each FLAIR ROI. The registered ROIs in the left hemisphere were flipped to the right hemisphere in the MNI space. All tumor ROIs in the MNI space were superimposed to create a voxel-wise frequency map of tumor occurrence in the hypometabolic and hypermetabolic groups separately, and used for the following analysis of differential involvement (ADIFFI) statistical mapping technique [9, 10].
The demographic (sex, age, WHO grade), molecular (IDH/1p19q) status, initial symptoms (seizure; focal neurological deficit, i.e. aphasia, hemiparesis, and muscle weakness; neurological symptoms, i.e. optic/olfactory/hearing/tasting abnormalities, dizziness, and vertigo; headache; other symptoms, i.e. mental/personality changes, unusual feelings; or incidental), and imaging metrics (nSUVmax, FLAIR volume, and BTV) were compared between the FDOPA hypometabolic and hypermetabolic LGGs using the Fisher’s exact or Mann–Whitney U test.
ADIFFI consisted of first constructing a 2 × 2 contingency table comparing two differential phenotypes (e.g. phenotypes A and B) and tumor versus non-tumor for each image voxel. Next, a two-tailed Fisher’s exact test was performed on a voxel-wise basis. According to the Fisher’s exact test, the probability of obtaining an observed pattern in the 2 × 2 contingency table is given by
where a is the frequency of tumor occurrence in a particular voxel for phenotype A; b the frequency of tumor occurrence in a particular voxel for phenotype B; c the frequency of no tumor occurring in a particular voxel for phenotype A; d the frequency of no tumor occurring in a particular voxel for phenotype B; n the total number of patients; and the exclamation point represents the factorial operation. To calculate the significance of the observed pattern, the contingency table corresponded to the total probability of observing a pattern in the contingency table as extreme or more extreme. Then, the p value was recalculated from each voxel for all cases in which the marginal totals were the same as the observed tables, and only for cases in which the arrangement was as extreme as the observed pattern. We performed this iteratively so that values were incremented to calculate a more extreme pattern, adding the previous p value in each image voxel each time until the most extreme pattern was achieved, which varied from voxel to voxel. The final p value represents the probability of observing the given pattern in the contingency table by chance. The p values less than .05 were considered significant. Additional details are presented in a previous publication .
For the cluster-based permutation correction outlined by Bullmore et al.,  a total of 500 random permutations were performed, the resulting ADIFFI-defined clusters with a connection of 18 directions were retained, and the 95% confidence interval (CI) for significant cluster size occurring by chance were documented. The cluster-size thresholds had a 5% probability of occurring by chance.
ADIFFI and cluster-based correction were additionally performed between different age, FLARI volume, or BTV with cut-off of the median value, and also among different molecular statuses.
Kaplan-Meier curves were used to depict differences in the OS, and the log-rank test was employed to compare OS between the FDOPA hypometabolic and hypermetabolic LGGs. Cox univariate regression analyses were conducted to investigate the association of OS with the age, nSUVmax, FLAIR volume, and BTV.
Statistical analyses were performed using MATLAB (R2019b; MathWorks, Natick, MA, USA) and GraphPad Prism (Version 8.3; GraphPad Software, La Jolla, CA, USA).