Next Generation Brain PET Imaging

Gold-standard quantitative imaging studies are often difficult to implement, limited by financial and logistical issues, or expose the patient to unnecessary risks. Deep learning has shown great promise in recent years for many medical applications; one use is to synthesize improved images. Such image transformation methods offer the potential to improve the quality, value, and accessibility of medical imaging. The goal of this project is to develop deep convolutional neural network approaches to FDG PET imaging, the most commonly performed clinical brain-focused PET study in the USA. Using simultaneous PET/MRI, we will train networks to produce diagnostic PET images from ultra-low dose PET and MR images. We will explore the three reimbursed clinical indications for this imaging modality (tumor recurrence, dementia, and epilepsy) using both quantitative metrics and repeated reader studies to assess equivalence and evaluate model generalization related to simultaneity, scanner type, age, sex, and disease prevalence. Next, we will evaluate whether we can move beyond ultra-low dose and remove the radiation dose altogether, synthesizing FDG brain PET images from MR inputs only, relying on the information in multi-modal functional MRI. Finally, we will assess whether we can use deep networks to combine imaging and non-imaging data such as clinical and genetic information to further improve image transformation and predict future images and image-based biomarkers. Significantly reducing or even eliminating the need for radiation to produce brain FDG PET images would be truly transformative while the ability to predict the future will enable personalized radiology and enhance our ability to perform clinical trials.