Virtual 18F-FDG Positron Emission Tomography Images Generated From Early Phase Images of 18F-FP-CIT Positron Emission Tomography Computed Tomography Using A Generative Adversarial Network in Patients with Suspected Parkinsonism

Abstract

Although attempts have been made to use the early phase images of brain N-(3-fluoropropyl)-2β-carboxymethoxy-3β-(4-iodophenyl) nortropane (18F-FP-CIT) positron emission tomography (PET)/computed tomography (CT) (FP-CIT) as a surrogate for 18F-fluorodeoxyglucose PET/CT (FDG) in patients with suspected parkinsonism, there has been no report of a direct comparison between the two. Also, artificial intelligence has not been applied to this particular area, while it has recently been applied to other medical imaging. This study aimed: 1) to generate virtual brain FDG images (FDGv) from early phase FP-CIT (FP-CITearly) images (FP-CITe) using a generative adversarial network (GAN) model in the DICOM format, and 2) to evaluate the similarity of FDGv compared to matched real FDG images (FDGr) for the possible use of FDGv as a surrogate for FDGr in patients with suspected parkinsonism. Methods: Brain FDG and dual-phase FP-CIT imaging obtained within 3 months of each other in 603 consecutive patients with clinically suspected parkinsonism between January 2013 and December 2019 were retrospectively reviewed. For deep learning, the training set included 452 randomly selected patients. Of the remaining 151 participants, 60 were again randomly selected and included in the testing set. A cycle GAN model based on unpaired image-to-image translation was used for training and testing. The FDGv were generated in the DICOM format. Mean standardized uptake values (SUVs) of brain regions, including the bilateral cerebral cortex, basal ganglia, thalamus, and whole cerebellum, were obtained, and SUV ratios (SUVRs) for each of these regions were calculated by dividing their respective mean SUV by the mean SUV of the whole cerebellum as a reference. Using Pearson correlation analysis and linear regression coefficients of SUVRs in each brain region obtained from FDGv, FP-CITe, and FDGr, comparisons between FDGv and FDGr (Comparison I), between FP-CITe and FDGr (Comparison II), and between FDGv and FP-CITe (Comparison III) were made. Results: There was no statistical difference in basic demographic characteristics between the training and test sets. Comparison II showed a considerably higher correlation than Comparison I. On the other hand, regression lines from Comparison I in the bilateral frontal, parietal, occipital, precuneus, and thalamus were closer to unity compared to those of Comparison II. Conclusion: A GAN-based image generator were able to produce FDGv from FP-CITe in the DICOM format. FDGv showed a somewhat lower correlation with FDGr than FP-CITe. On the other hand, the actual values of SUVRs and regression coefficients obtained from FDGv were closer to unity than those obtained from FP-CITe when correlating with FDGr. Therefore, FDGv may have the potential as a surrogate for FDGr with further optimization of deep learning. Thereafter, they may be helpful for reducing inconvenience, cost, and extra radiation exposure resulting from performing both FDG and FP-CIT imaging.