Silhouette-Based Markerless Motion Estimation of Awake Rodents in PET

The ability to image the brain of a freely moving rodent using motion-compensated PET presents many exciting possibilities for exploring the links between brain function and behavior. Markerless optical approaches for pose estimation have several potential advantages over marker-based methods: improved accuracy and increased range of detectable motion; no `decoupling' of marker and head motion; and no acclimatization of the animals to attached markers. Our aim in this work was to describe and validate a silhouette-based multi-camera method for estimating the pose of a rat. Random-walk and K-means clustering approaches were very adaptable to uneven lighting and generally provided excellent object segmentations. In obtaining a high quality rat model, shape-from-silhouette and laser scanning both resulted in useful models; laser scanning provided sub-millimeter resolution with very few artifacts and was the method of choice. In our experimental validation, the 3D-2D (model-silhouette) optimization clearly converged to sub-degree and sub-millimeter alignment of the measured and estimated silhouettes. The average discrepancy between test points transformed using the estimated versus ground-truth poses was 0.94 mm ± 0.51 mm. This investigation focused on rigid motion of a rat phantom as a proof-of-principle of the technique. Future work will focus on investigating the potential of designing a non-rigid rodent body model in order to apply the method to a freely moving animal during PET imaging.