Synthetic Full Dose Cardiac PET Images from Low Dose Scans Using Conditional GANs

Positron emission tomography (PET) is often used to characterize diagnosing cancer, heart diseases, and neurological disorders. Radiation exposure to patients results from the injection of radioactive tracers. Researchers have proposed deploying various artificial intelligence (AI) methods to enhance image quality while reducing radiation dose. Lowering the dose, however, will result in low signal-to-noise-ratio (SNR) and loss of information, thus leading to difficulties with clinical decision making. To address this problem, we evaluated the performance and generalizability of a deep-learning algorithm for improving the quality and accuracy of low dose cardiac PET image quality. Our model structure is based on a conditional generative adversarial neural network (cGAN) including a U-net generator with skip connections and a discriminator. Model training and validations were performed on a low dose PET dataset simulated from a full dose dataset acquired from 40 patient. Low dose cardiac 18F-fluorodeoxyglucose (
¹⁸
F-FDG) PET scans were reconstructed using reduced PET raw data by randomly selecting 5 to 14% of the total events in the PET list mode data. We demonstrate through quantification that the cGAN framework successfully creates synthetic full dose images from low dose scans while preserving the original image resolution with significant noise correction.