Deep Learning Aided Motion Correction for Low-Dose Brain PET

We have previously developed a deep learning (DL) aided PET motion correction method which successfully detected and corrected for head motion in the situation of a rapidly changing tracer distribution such as the initial tracer uptake period – immediately after injection. Here we extend this data driven method to a general low-dose scenario where tracer distribution is relatively constant but available counts are very low, such as for low-dose imaging. Data from four tracers were used ([
¹¹
C]-RAC, [
¹⁸
F]-FDG, [
¹¹
C]-PBR, and [
¹¹
C]-DTBZ) to form a representative dataset for training and testing of a model capable of mapping low-count, high temporal resolution frames (acquired during a low-count period of the scan), to higher-count frames containing better structural features than the original low-count frames. The ability to recover a known motion trace that was artificially injected into low-count subject data was tested. The use of the DL model reduced the standard deviation of the motion estimation error from 0.5 mm and 0.8 degrees (when registration was performed without the aid of DL) to 0.3 mm and 0.5 degrees. These results were then validated with real human motion using the estimated motion from simultaneous fMRI acquisition as the ground truth.