Pseudo-normal PET Synthesis with Generative Adversarial Networks for Localising Hypometabolism in Epilepsies

Abstract

[${}^{18}$F]fluorodeoxyglucose (FDG) positron emission tomography (PET) aids in the localisation of the epileptogenic zone in patients with focal epilepsy, especially when magnetic resonance imaging (MRI) is normal or non-contributory. We propose a two-stage deep learning framework to support the clinical evaluation of patients with focal epilepsy by identifying candidate regions of hypometabolism in [¹⁸F]FDG PET scans. In the first stage, we train a generative adversarial network (GAN) to learn the mapping between healthy [¹⁸F]FDG PET and T1-weighted (T1w) MRI data. In the second stage, we synthesise pseudo-normal PET images from T1w MRI scans of patients with epilepsy to compare to the real PET scans. Comparing the estimated pseudo-PET images to the true PET scans in healthy control data, our GAN produced whole-brain mean absolute errors of $0.053 \pm 0.015$, outperforming a U-Net ($0.058 \pm 0.021$) and a high-resolution dilated convolutional neural network ($0.060 \pm 0.024$; all images scaled 0–1). In a sample of 20 epilepsy patients, we created Z-statistic images (with thresholding at +2.33) by subtracting the patient’s true PET scans from their estimated pseudo-normal PET images to identify regions of hypometabolism. Excellent sensitivity for lobar location of abnormalities ($92.9 \pm 13.1\%$) was observed for the seven cases with MR-visible epileptogenic lesions. For the 13 cases with non-contributory MR, a lower sensitivity of $74.8 \pm 32.3\%$ was observed. Our method performed better than a statistical parametric mapping analysis. Our results highlight the potential of deep learning-based pseudo-normal [¹⁸F]FDG PET synthesis to contribute to the management of epilepsy.