Patients
We retrospectively selected newly-diagnosed glioma patients for whom 18F-FDopa PET had been performed as part of the initial tumour characterisation, in the Department of Nuclear Medicine at the CHRU of Nancy, between February 2018 and June 2020. The final selection included: (i) patients with a neuropathological diagnosis based on the WHO 2016 classification [6] and with a maximum delay from the time of the 18F-FDopa PET of 150 days for diffuse grade II or III gliomas and 60 days for glioblastomas [2,4], (ii) patients with available raw-data for a-posteriori reconstruction and, (iii) patients with a visually abnormal 18F-FDopa uptake, i.e. by excluding isometabolic and photopenic gliomas [16,17]. The institutional ethics committee (Comité d’Ethique du CHRU de Nancy - FRANCE) approved the evaluation of retrospective patient data on August 26, 2020. The trial was registered at ClinicalTrials.gov (NCT04469244). This research complied with the principles of the Declaration of Helsinki. Informed consent was obtained from all individuals included in the study.
18F-FDopa PET Acquisition and Image Reconstruction
Patients were instructed to fast for at least 4 hours prior to the examination, and were pre-medicated with Carbidopa one hour prior to the examination to increase tracer uptake in the brain [13]. Patients were scanned with a digital Vereos PET/CT camera (Vereos, Philips, Cleveland, Ohio). Immediately after recording CT images (100 kV, 80 mAs), a 30-min 3D list-mode PET acquisition was initiated concomitantly to the bolus injection of 2MBq of 18F-FDOPA per kilogram of body weight. Static PET images were reconstructed from the list mode data acquired 10 to 30 min post-injection [8,18]. Dynamic images were reconstructed using two different temporal sampling protocols depending on the dynamic analyses carried out. Where an input function was required, the temporal sampling protocol was similar to what was recommended in the EANM/SNMMI guidelines, i.e. 12 frames of 5 s, 6 frames of 10 s, 6 frames of 30 s, 5 frames of 60 s and 4 frames of 300 s [8]. The temporal sampling protocol consisted of 30 x 60 s frames for the models not requiring any input function [19].
Static images were reconstructed using the time-of-flight information and a high-resolution protocol with the OSEM 3D algorithm (2 iterations, 10 subsets, a deconvolution of the point spread function (PSF) and 256 x 256 x 164 voxels of 1 x 1 x 1 mm3) while a protocol with a lower spatial resolution was used for dynamic images to limit the level of noise, i.e. 3 iterations, 15 subsets, without PSF and 128 x 128 x 82 voxels of 2 x 2 x 2 mm3 [20].
All images were corrected for attenuation using CT, dead time, random and scattered coincidences during the reconstruction process.
Segmentation
On static images, the healthy brain uptake was initially measured using a merged volume of interest (VOI) consisting of a crescent-shaped region of interest manually positioned on three consecutive slices of the unaffected hemisphere so as to comprise both white and gray matter, while the tumour VOI was segmented semi-automatically using a threshold of 1.6 healthy brain SUVmean as previously recommended [19,21,22]. The arterial input function VOI was subsequently placed into the internal carotid using initial dynamic frames to identify early vascular phases [23].
All volumes of interest were segmented using the LifeX software (lifexsoft.org) [24] and were visually inspected by an experienced physician (A.V.) to ensure the quality of the methods applied. Healthy brain was considered as the reference region due to its non-specific uptake, where required [21].
Extraction of time-activity curves
For dynamic images reconstructed using the protocol with 30 x 60 s frames, each dynamic frame was first registered to the associated CT image, in order to correct for any potential patient movement during the acquisition [25]. These transformations, representing the evolution of the patient’s movements over time, were interpolated to the time frames of the other protocol for models involving an input function. Indeed, the first frames from images reconstructed with models involving an input function are very short and suffer from noise, which makes the registration very challenging.
Blood and brain time-activity curves (TACs) were extracted by retrieving the SUVmean for each frame in their respective VOI. Tumour TACs were computed by retrieving the SUVmean for each frame in the volume corresponding to the SUVpeak of the tumour VOI on the static image, to represent the most aggressive part of the tumour [19].
Input Function Pre-processing
Since no arterial blood sampling was performed in this study, an image-derived input function was used for analyses that required one. TACs representing the evolution of the arterial blood activity were obtained from internal carotid VOIs, and were fitted using linear interpolation to the peak followed by a tri-exponential function after the peak [26]. In the case of 18F-FDOPA, the plasma 18F-FDOPA TAC is obtained after correcting for OMFD as well as other METS created in the peripheral tissues [15]. This plasma 18F-FDOPA TAC can be obtained from the blood TAC if the haematocrit level and the proportion of each 18F-labelled entity in the respective plasma TACs are known [14]. These values were retrieved from the literature, specifically from Huang et al. [15] who used a haematocrit of 40% and from Melega et al. [27] who reported the metabolite proportions for patients pre-medicated with 100 mg of Carbidopa which is identical to the pre-medication schedule of our patients. To extrapolate the proportion of metabolites at any time, the measured fractions of plasma radioactivity were fitted using the following equations for the plasma fractions of DOPA, OMFD, and METS respectively:
![](https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1627272172.png)
where t is the time in minutes.
Dynamic models
The pre-processing steps and input data for each dynamic model are presented in Figure 1.
Semi-quantitative models
To overcome noise effects, tumour TACs were first fitted using non-linear least square optimisations and a specific tumour vascularisation function (patent WO/2008/053268, entitled "Method and System for Quantification of Tumoral Vascularization") [4,19,28]. Semi quantitative (SQ) model parameters, time-to-peak (TTP) and slope, were respectively computed as the time from the beginning of the dynamic acquisition to the maximum uptake value and as the slope of the linear regression of the data between the 10th and 30th minute [4]. The reference semi quantitative (Ref SQ) model was conducted as an assessment of other studies where tumor-to-normal brain ratio dynamic values were used to overcome the Carbidopa effect [4,19,28] even though such normalisation was not needed in this study. For this purpose, healthy brain TACs were also fitted similarly to tumour TACs, and TACsratio representing the evolution of the ratio between tumour and brain fitted TACs were computed. TTPratio and sloperatio were computed from the TACsratio similarly to what was done for tumour TACs.
Graphical models
Among all graphical models available, the Logan graphical model [29] with the computation of the equilibrium volumes of distribution is particularly suited to 18F-FDOPA in gliomas since there is no evidence that 18F-FDOPA is trapped in tumours [13]. The Logan graphical model (Logan) was performed with the slope computed between 15 and 30 min p.i. as previously suggested [13]. The equilibrium volumes of distribution Ved and IntLogan were computed respectively as the slope and the intercept of the graphical model. This model was also performed using a healthy brain reference region (Ref Logan) as it can be done with the Patlak graphical model in 18F-FDOPA PET imaging in parkinsonian syndromes [30]. Two parameters were also extracted with regression between 15 and 30 min, namely the distribution volume ratio (DVR) and the relative residence time (RRT) computed respectively as the slope and the intercept of the graphical model with the reference region.
Compartmental model
The model used was a simplified two-tissue compartmental model (2TCM) adapted from the original publication of Huang et al. [15] and already validated for compartmental analysis of 18F-FDOPA imaging in gliomas [14]. Four rate constants (K1, k2, k3, k4) as well as the blood volume fraction (Vb) were estimated by fitting the two-tissue compartmental model to tumour TACs. The net influx rate constant Ki was computed from the previously estimated four rate constants as
![](https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1627272293.png)
Statistical Analysis
Categorical variables are expressed as numbers and percentages and continuous variables as medians [first quartile; third quartile] because variables did not follow a normal distribution. Intergroup comparisons were performed with the Chi-squared test for categorical variables and the Mann-Whitney test for continuous variables.
For the overall comparison of the different kinetic models, parameters belonging to the same model were used to construct a multivariate model. This multivariate model was a general linear model with variables selected automatically in a stepwise manner with both forward and backward selection minimising the Akaike Information Criterion. Comparisons of performance of the final models were carried out with the one-sided comparison of superiority pairwise Delong tests [31].
The ability of each individual extracted parameter to predict an IDH mutation was assessed using receiver operating characteristic (ROC) curves from which area under the curve (AUC), sensitivity, specificity and accuracy were computed. The optimal threshold was determined by selecting the point on the curve closest to (0,1). Spearman coefficients were calculated to assess correlations between each extracted parameter of the different models.
p values were adjusted using the Benjamini-Hochberg correction and p values lower than 0.05 were considered significant. Analyses were performed with the R software version 3.6.2 (R Foundation for Statistical Computing, Vienna, Austria).