We retrospectively analyzed the patients who received consecutive ¹⁸F-FET PET/CT brain imaging at the PET Center of Huashan Hospital, Fudan University from November 2017 to February 2019. A total of 58 untreated adult glioma patients were identified based on inclusion and exclusion criteria Supplementary S1 . Baseline clinical and demographic characteristics including age, gender, tumor location, WHO grade, and IDH mutation status, were derived from medical file system of Huashan Hospital, Fudan University.

All patients had fasted for at least 4 hours before imaging. 20 minutes after an intravenous bonus injection of ¹⁸F-FET (370 ± 30 MBq), a static PET scan was collected for 20 minutes with a Siemens Biograph 64 HD PET/CT (Siemens, Erlangen, Germany) in 3-dimensional (3D) mode for 50 patients in this cohort. 8 patients underwent dynamic ¹⁸F-FET PET scanning lasted longer than 40 minutes after the tracer injection; 20–40 minutes images were reconstructed for diagnosis and analysis. Attenuation correction was performed using a low-dose CT (tube current 150mAs, voltage 120kV, Acq. 64*0.6mm, convolution kernel H30s, slice thickness 5mm, interslice gap 1.5mm) before the emission scan. After acquisition, the PET images were reconstructed by filtered back projection algorithm with Gaussian filter and full-width-at-half-maximum at the center of the field of view of 3.5mm). The reconstructed brain PET image matrix size was 168*168 with voxel size of 2.04*2.04mm. And the reconstructed brain CT image matrix size was 512*512 with voxel size of 0.59*0.59 mm.

Histological specimens were obtained by surgical resection or stereotactic brain biopsy. H&E staining and immuno-histochemical analysis were performed by experienced neuropathologists according to the 2016 WHO classification. IDH genotype status was identified with an antibody to the IDH1 (R132H) mutation by immunohistochemical staining. For 31 out of the 58 patients, IDH mutational status from the immuno-histochemical staining was further confirmed by sequencing.

PET/CT imaging was first analyzed using a dedicated workstation (Siemens Syngo.via) to obtain semi-quantitative parameters in 3D volumes. Background mean standardized uptake value (SUV) was measured initially in a crescent-shape area, including both gray and white matter on the contralateral hemisphere as mentioned (27). The brain structural MRI of patients were reviewed initially to locate the tumor region. A predefined threshold of 1.6-times of the background mean SUV was applied for tumor volume of interest (VOI) delineation to derive lesion maximal SUV (SUVmax), peak SUV (SUVpeak), mean SUV (SUVmean), the standard deviation of lesion standardized uptake value (SUV_SD), metabolic tumor volume (MTV) and total lesion tracer uptake (TLU) (28). Maximal tumor to background ratio (TBRmax), peak tumor to background ratio (TBRpeak), and mean tumor to background ratio (TBRmean) were calculated by the division of tumor VOI SUVmax, SUVpeak and SUVmean with background SUVmean. TLU was defined as the MTV multiplied by the tumor lesion SUVmean. For those multifocal glioma patients in our group, the specific surgical resected or biopsied lesion for pathological examination were included in our research to avoid bias.

Two experienced nuclear medicine physicians (WY Zhou: 5 years’ experience, T Hua: 10 years’ experience) performed lesion delineation and obtained two sets of semi-quantitative parameters separately. Brain midline involvement status were judged respectively by the two physicians after lesion delineation based on ¹⁸F-FET PET imaging. The brain midline structures included corpus callosum, cingulate gyrus, bilateral thalamus, third ventricle, brain stem and cerebellar vermis. The inter-observer agreement indices were evaluated for those obtained semi-quantitative parameters and glioma location status.

Univariate logistic regression was applied to filter the conventional semi-quantitative parameters for further model construction. The best-performance semi-quantitative parameter was combined with dichotomous brain midline involvement status for simple predictive model building.

The original ¹⁸F-FET PET/CT data in digital imaging and communications in medicine (DICOM) format were converted into NIFTI format using SPM12 (Welcome Trust Centre for Neuroimaging, London, UK; http://www.fil.ion.ucl.ac.uk/spm).Then the tumor lesions were manually delineated on ¹⁸F-FET PET images using the ITK-SNAP software (http://www.itksnap.org/pmwiki/pmwiki.php) with the reference to patients’ structural MRI by the above-mentioned two physicians separately. If the divergence of a patient’s segmentation by the two physicians was less than 5%, the final segmentation of this specific patient was determined as the overlapping region of the two separate VOIs, and if the divergence was more than 5%, a senior nuclear medicine physician with over 25 years’ experience (YH Guan) made the final decision. After that the masks of those final tumor VOIs delineation on the ¹⁸F-FET PET images were applied to the co-registered CT images for the CT lesion delineation obtaining.

The registered ¹⁸F-FET PET/CT images and the VOIs were used to extract features. Prior to feature extraction, image standardization was implemented as follows: sitkBSpline interpolation resampling techniques were used to standardize the image scale in the slice, resulting in a pixel size of 2.5mm×2.5mm×2.5mm for PET and 1.5mm×1.5mm×1.5mm for CT, respectively. Two sets (105*2) of radiomics features from VOI lesions of the original PET and CT images were extracted with no further transformation or filtering using PyRadiomics (https://github.com/Radiomics/pyradiomics) (29), including 13 shape and size features, 18 first-order features and 74 texture features. The shape and size features included descriptors of the 3D size and shape of the VOI, which are independent from the gray level intensity distribution in the VOI. First-order features included the maximum, mean, and average absolute deviation of the gray-level intensity values in the VOIs. The texture features consisted of second-order features and are used to express heterogeneity in the tumor, from common matrices such as gray-level co-occurrence matrix (GLCM), gray-level size zone matrix (GLSZM), gray-level run length matrix (GLRLM), gray level dependence matrix (GLDM) and neighboring gray tone difference matrix (NGTDM). The feature extraction algorithms were standardized by referring to the Image Biomarker Standardization Initiative (IBSI) (30). Considering the relative low-resolution nature of PET images, radiomics feature extraction was conducted only on the original PET images with no further transformation or filtering. The details of extracted radiomics features were presented in Supplementary S2 .

Owing to the relatively small number of glioma patients in our cohort, less than five radiomics features were filtered for predictive model construction to minimize overfitting or selection bias in our radiomics models (13, 31). Least absolute shrinkage and selection operator (LASSO) and all-subsets regression were performed to filter the most significant radiomics features combinations for IDH genotyping. The LASSO algorithm adds a L1 regularization term to a least square algorithm to avoid overfitting, which is suitable for the regression of high dimensional data (32). Logistic regression analysis was then utilized to integrate the filtered features after further all-subsets regression selection. Calculated-score of three predictive models originating from radiomics features of PET, CT and their combined modalities were obtained by the linear fusion of the selected discriminating radiomics features weighted by their respective coefficients.

The performance of the predictive models was evaluated by the receiver-operating characteristic (ROC) analysis and compared by the DeLong test. Additionally, the area under the ROC curve (AUC) with 95% confidence interval (CI), sensitivity, specificity, accuracy, positive predictive value and negative predictive value were calculated for each predictive model.

The four models for IDH genotype prediction were evaluated with k (k=3,5,10) fold cross-validation. The clinical application value of the predictive models was determined and compared through the decision curve analysis (DCA) by quantifying the net benefit to the patient under different threshold probabilities in the cohort (33).

Descriptive statistics were presented as mean ± standard deviation or median and range. Categorical variables were expressed as percentages. An independent sample t-test was used to compare two groups, while the chi-square test was applied to calculate p values for categorical variables. The Mann-Whitney rank sum test was used when variables were not normally distributed. Inter-observer agreements on ¹⁸F-FET PET metrics and dichotomized location results were assessed with interclass correlation coefficients (ICC) and Cohen’s kappa coefficient analysis respectively, defined as poor (less than 0.2), fair (0.21-0.4), moderate (0.41-0.6), good (0.61-0.8), and very good (0.8-1.0). Univariate and multivariate logistic regressions were used to identify the predictive factors for IDH mutation. The package “glmnet” was used to perform LASSO binary logistic regression analysis, and the “leaps” package to achieve all-subset regression algorithms, and the “bootstrap” package, to perform cross-validations by the R software, version 4.0.4 (http://www.r-project.org/). The ROC analyses and the DeLong test were performed by MedCalc for windows (version11.3.3.0, MedCalc software, Mariakierke, Belgium). All other statistical analysis was performed with Prism Software version 8.0 (GraphPad, San Diego. CA). In all analyses, p < 0.05 was considered to indicate statistical significance.

Demographic and clinical data of the 58 enrolled glioma patients were summarized in Table 1 . The patient recruit process is presented in Figure 1 . The time interval between PET imaging and subsequent tumor resection or biopsy was no more than 90 days for grade II or III patients and not exceeding 30 days for grade IV glioblastomas. A detailed chronologically enrolled patients’ clinical data was provided in Supplementary S3 .

The dichotomized brain midline involvement status of the interested tumor lesion yielded unified results for two separate readers, and the inter-observer kappa was satisfactory (κ=1.0, p<0.0001). The ICC also showed very good agreement for the ¹⁸F-FET PET volume-based semi-quantitative parameters (ICC>0.95, p<0.0001). Therefore, the results of reader one (WY Zhou) was applied for further analysis. The absolute values for all ¹⁸F-FET PET metrics based on different IDH-genotype were also provided in Table 1 .

Significant difference of SUV_SD (0.2145 ± 0.1577 vs. 0.3782 ± 0.3158, p = 0.034) and TLU [12.23(2.82-25.33) vs. 26.57(11.97-66.08), p = 0.046] could be observed in the two groups with different IDH genotype. While MTV of ¹⁸F-FET PET uptake was insignificant (p=0.239), the P values of TBRmax, TBRmean and TBRpeak were around the borderline (p=0.060, 0.053, 0.067, respectively). Further univariate logistic analysis results indicated that SUV_SD and the brain midline involvement status were significant independent predictors for IDH mutation. SUV_SD [≤0.23 vs. >0.23, odd ratio (OR): 3.208, 95%CI (1.013-10.163), p = 0.048] and brain midline involvement [no vs. yes incidence, OR: 9.714, 95%CI (2.412-39.125), p = 0.001] were associated with a higher of IDH mutation. Considering the collinearity of SUV_SD and TBRs (variance inflation factor threshold <5), SUV_SD and dichotomized tumor location status were selected to build the simple generalized linear model with reasonable IDH genotype predictive performance [AUC (95%CI) = 0.786 (0.659-0.883), sensitivity = 85.0%, specificity = 71.1%, accuracy = 75.9%]. The predictive scores of simple model for each patient were calculated using the following formula: 0.7836-2.1718*location (midline structure involved: 1, uninvolved: 0)-2.1462*SUV_SD. IDH-mutants had higher calculated-scores than IDH-wildtypes in the simple model (-0.0025 ± 0.9398 vs. -1.3997 ± 1.2489, p < 0.001).

Three radiomics predictive models were developed by machine-learning based algorithms. Three PET radiomics features were selected for the building of the PET-Rad Model. Three CT radiomics features were selected for the building of the CT-Rad Model. Three CT radiomics features and one PET radiomics feature were filtered for the building of the PET/CT-Rad Model (shown in Figure 2 ). The relative importance analysis of the PET/CT-Rad Model indicated one of the three CT radiomics features (CT_glcm_InverseVariance) had the most important weight with 30.25% and the PET radiomics feature (PET_glcm_JointEnergy) had the third weight out of the four features (26.41%) in the PET/CT-Rad Model (shown in Figure 3 ). The calculated-scores of the three radiomics models for each patient were obtained using related formulas and shown as follows:

Results showed IDH-mutants had higher calculated-scores than IDH-wildtypes in the PET-Rad, CT-Rad and combined PET/CT-Rad models (1.1262 ± 1.8048 vs. -2.6648 ± 2.1637, p < 0.001; 1.0008 ± .2.0396 vs. -2.5214 ± 2.5610, p < 0.001; 0.1536 ± 1.2660vs. -1.6186 ± 1.6268, p < 0.001; respectively) ( Figures 4A , B ).

ROC analysis results indicated the AUC of the PET-Rad Model was 0.812 (95% CI 0.688-0.902), with a sensitivity of 80.0%, a specificity of 73.7% and an accuracy of 75.9%. The AUC of the CT-Rad Model was 0.883 (95% CI 0.771-0.952), with a sensitivity of 85.0%, a specificity of 76.3% and an accuracy of 79.3%. The PET/CT-Rad Model achieved the highest AUC (0.912, 95%CI 0.808-0.970), with a sensitivity of 85.0%, a specificity of 86.8% and an accuracy of 86.2% (Details were provided in Table 2 ). DeLong test results indicated that PET/CT-Rad Model outperformed the PET-Rad Model and the simple model in IDH genotype prediction (p = 0.048 and p = 0.034, respectively) ( Figure 5 ). No statistically significant difference could be observed among the PET-Rad Model, CT-Rad Model and simple model.

The decision curve analysis results revealed that the PET/CT-Rad Model was the most satisfactory predictive model in differentiating IDH mutation status in these four models (shown in Figure 6 ). Representative glioma patients of different IDH genotypes were provided in Figure 7 .

We further explored the combination of the dichotomized tumor location status and PET/CT-Rad Model for IDH mutation status prediction, and results indicated the predictive efficacy slightly enhanced in terms of AUC and accuracy (from 0.912 to 0.917 and 86.2% to 87.9%, respectively), although the enhancement was statistically insignificant.