Multi-tracer Deep Learning for PET Head Motion Correction

Head motion during brain positron emission tomography (PET) image acquisition decreases image quality and increases quantification errors. So far, no robust approach allows continuous head motion tracking without using an external device. In this feasibility study, we propose a multi-tracer deep learning head motion correction (mDL-HMC) model to predict rigid head motion based on supervised learning of gold-standard camera motion tracking. mDL-HMC utilizes a novel feature-wise transformation block to condition the network to different tracer types. Quantitative and qualitative results from five subjects demonstrate that training this model with combined data from two different tracers (FDG and UCB-J) improves motion correction performance compared to single-tracer models. Our head motion correction results have the potential to eliminate the need for external motion tracking hardware.