ABSTRACT
Head motion during brain PET imaging is not uncommon and can negatively affect image quality. Motion correction techniques typically either use hardware to prospectively measure head motion, or they divide the acquisition into short fixed-frames and then align and combine these to produce a motion free image. The aim of this work was to retrospectively detect when motion occurred in PET data without the use of motion detection hardware, and then align the frames defined by these motion occurrences. We describe two methods that use either principal component analysis or the motion induced spatial displacements over time to detect motion in raw time-of-flight PET data. The points in time of motion then define the temporal boundaries of frames which are reconstructed without attenuation correction, aligned and combined. Phantom and [18F]-Fallypride patient acquisitions were used to validate and evaluate these approaches, which were compared with motion estimation using 60 s fixed-frames. Both methods identified all motion occurrences in phantom data, and unlike the fixed-frame approach did not exhibit intra-frame motion. With patient acquisitions, images corrected with the motion detection methods increased the average image sharpness by the same amount as the fixed-frame approach, but reduced the number of reconstructions and registrations by a factor of 3.4 on average. Detecting head motion in raw PET data alone is possible, allowing retrospective motion estimation of any listmode brain PET acquisition without additional hardware, subsequently decreasing data processing and potentially reducing intra-frame motion.
