摘要
Background: Head motion impairs brain PET resolution and quantification. On the NeuroEXPLORER (NX) (1), United Imaging markerless motion tracking (UMT) (2) can fail with poor face coverage or facial expressions (3). We extended our supervised method (4) to unsupervised deep learning for head motion estimation model (uDL-HME) for multi-tracer motion estimation.
Methods: NX human scans included [11C]raclopride (n = 8), [11C]UCB-J (n = 18), and [18F]FE-PE2I (n = 10), with two scans per tracer held out. Five-second non-attenuation-corrected YRT-PET (5) reconstructions underwent truncated singular value decomposition of the space–time matrix to suppress noise and improve keypoint extraction for KeyMorph (6) rigid motion estimation. Training used [11C]UCB-J dataset and a normalized cross-correlation loss, followed by adaptation to other tracers. Trajectories were compared with objectively quality-controlled UMT (7) in six scans. For motion correction, vendor software used estimated poses to identify motion-free frames (MFFs), reconstruct MFFs at different head positions separately, and combine them into predefined frames. In six test scans, we compared uDL-HME with post-reconstruction registration (PRR) using time-activity curve (TAC) variability: cubic-fit residual standard deviation normalized to mean activity (SD%). One motion-affected [18F]FE-PE2I scan underwent SRTM2, including target/reference TAC exchanges.
Results: Rotation and translation differences from UMT were 0.56 ± 0.26° and 0.69 ± 0.31 mm. TAC variability was lower with uDL-HME than PRR for all tracers; overall average 2.01% vs 3.14% (Fig. 1a). In the motion-affected scan, SD% improved from 3.1–7.9% to 0.9–2.3% (Fig. 1b), while BPND ranges were 0.40–1.24 (PRR) and 0.38–1.23 (uDL-HME). PRR TACs were generally lower across regions. Replacing PRR target TACs increased BPND by 0.04–0.13; subsequently replacing the reference decreased it by 0.09–0.14 (Fig. 1c).
Conclusion: uDL-HME enables label-free, data-driven motion estimation with submillimeter mean translational differences from UMT and smoother TACs than PRR across three tracers. Opposing target/reference effects can mask motion-related kinetic changes.
