Deep-Learning to generate Standard-Count PET from Low-Count PET

Project Description

Low-count PET (LC-PET) acquisition in preclinical context has multiple benefits in terms of improving animal logistics, preserving integrity in longitudinal studies and in increasing scanner throughput while decreasing radiation risk. Low count imaging implemented either by reducing scan-time or injected radiolabeled agent can lead to significant reduction in photon count level thus generating PET images with poor diagnostic quality, quantitative uncertainties and detection difficulties. To address this tradeoff challenge our aim is to develop a deep-learning based novel framework for generating Standard-Count PET (PET) images from LC-PET across different photon levels optimized for preclinical standards. The quality of the DL generated SC-PET images would be evaluated using not only traditional fidelity based image quality metrics but also in context of task-based segmentation performance and task-based quantification analysis (SUV Measurements and Radiomics) with respect to true SC-PET. This would validate the diagnostic ability and quantitative performance of the DL generated SC-PET being used for analyzing therapeutic efficacy.