An Evaluation of a Deep learning approach for Radiation Dose Reduction in 18F-FDG PET/MRI Pediatric Epilepsy Imaging

Epilepsy is a degenerative brain disease characterized by abruption of neural activities that result in seizures. The onset of epileptic seizures are usually from a primary source - the epileptogenic foci (EF) which could be distributed to nearby neurons and tissues. Accurate localization of EF is critical in epilepsy cases where drug treatment has failed, and surgery is indicated to resect the EF to alleviate seizure. Typically, hybrid positron emission tomography (PET) and computed tomography (CT) imaging are performed to functionally localize the EF in drug-resistant epilepsy for surgical planning when anatomical abnormalities representing the EF cannot be identified on magnetic resonance imaging (MRI). The recent introduction of integrated PET/MRI scanning has significantly enhanced the localization of EF and eliminated the use of CT for PET attenuation correction (AC), minimizing radiation exposure particularly in radiosensitive patients such as pediatrics. The objective of this thesis is to develop an image analysis approach to further reduce PET radiation dose for optimal pediatric epilepsy PET/MRI. First, to eliminate CT for PET AC, a robust deep learning approach validated in pediatric populations for synthesizing CT from MRI was implemented after performing a rigorous systematic review and meta-analysis of state-of-the art machine learning (ML) AC methods. Next, a deep learning tool using the Self-SiMilARiTy-Aware Generative Adversarial framework (SMART) was developed and evaluated for denoising of PET images acquired with 90% reduction in PET dose, to generate high quality PET images. By combining ML-AC and SMART-PET, this work proposed an approach to drastically reduce the radiation exposure for high quality pediatric epilepsy PET imaging from ~6 mSV in PET/CT to ~0.5 mSV in low-dose PET/MRI.