Prediction of FLAIR MRI from 18F-FDG PET/CT for the Evaluation of White Matter Hyperintensity Using Generative Adversarial Network

Abstract

White matter hyperintensities (WMH) may decrease cortical glucose metabolism and appear hypodense on F-18-fluorodeoxyglucose (FDG) positron emission tomography (PET) and computed tomography (CT), respectively. Currently, T2-weighted fluid-attenuated inversion recovery (FLAIR) images on magnetic resonance imaging (MRI) are considered as a sequence of choice to evaluate WMH. This study aimed to use a generative adversarial network (GAN) to predict FLAIR MRI images from F-18-FDG PET/CT. From 2017 to 2019, we selected 167 patients who had MRI and FDG PET/CT scans within 6 months. We categorized WMH into three groups using the Fazekas scale and trained a GAN model to predict MR FLAIR images from PET and CT data (pix2pix-PT), or only CT data (pix2pix-CT). We compared these predicted images with actual MR FLAIR images, then performed WMH segmentation and volume estimation, assessing their agreement. To predict ground-truth FLAIR images, the pix2pix-PT method demonstrated superior performance compared with pix2pix-CT, as evidenced by the lower NMAE and higher PSNR in all groups. Integrating these findings with the segmentation results, the performance of the pix2pix-PT model in WMH segmentation was differential across groups. Notably, the pix2pix-PT model accurately segmented WMH lesions over 0.3 cm(2) without false positives or negatives and maintained a DSC above 0.7 for lesions over 2 cm(2). For WMH volume estimation, the pix2pix-PT method showed excellent correlations in Group 2 (r = 0.903) and Group 3 (r = 0.984), and moderate in Group 1 (r = 0.780). In this study, the prediction of T2-weighted FLAIR MR images using the GAN model was better achieved when both FDG PET and CT data were provided to the model, compared to CT data alone. Predicted T2-FLAIR images derived from our model could aid in selecting patients who need MRI to assess WMH and related vascular pathology.