Super-resolution in brain PET Using a Real Time Motion Capture System

Objectives: Super-resolution (SR) is a powerful methodology seeking to improve image resolution by exploiting the increased spatial sampling information obtained from multiple acquisitions of the same object performed with accurately known sub-resolution shifts [1]. The goal of this work is to develop and evaluate a SR estimation framework for brain positron emission tomography (PET), taking advantage of a high-resolution infra-red tracking camera to measure object shifts precisely and continuously.

Methods: Moving phantom and non-human primate (NHP) experiments were performed on a PET/CT scanner (Discovery MI, GE Healthcare) using an external motion tracking device (Polaris Vega, NDI). The Polaris Vega, which yields a tracking accuracy up to 0.12 mm isotropic and a 60 Hz frame rate, was integrated with the scanner through a custom-built interface sending trigger events synchronized with the motion capture frames to the list mode. A transformation linking the Polaris and scanner coordinates space was determined by six paired measurements of high-resolution CT scans (0.7x0.7x0.6 mm) and Polaris tracking of the same reference object by a least squares minimization to spatially align the two instruments. To enable SR, a list-mode Ordered Subset Expectation Maximization (OSEM) PET reconstruction algorithm was developed, incorporating the high-resolution tracking data from the Polaris Vega to correct motion for measured line of responses (LORs) on an event-by-event basis. A Hoffman phantom, filled with 3mCi 18F, was scanned for 15 min in list mode while undergoing continuous rotation/translation movements introduced by a QUASAR system (Modus QA). Similarly, a rhesus monkey administered with 11 mCi 18F-FDG was scanned for 15-min in list mode with continuous head motion induced manually. For both scans, motion was tracked at all time using the Vega and reflective markers rigidly attached to the targets. Reference static PET acquisitions were also performed without inducing movement. PET images were reconstructed with 3 different methods: (A) OSEM algorithm applied to the static reference scan data (2 mm voxel size), (B) OSEM algorithm applied to the static scan on smaller voxel size (1 mm), (C) proposed SR algorithm applied to the moving scan with 1 mm voxel size. Attenuation and sensitivity effects were accounted for during reconstruction. Iteration numbers were chosen to obtain matched image noise levels for method (A) and (C). For method (B), we used the same number of iterations as for (C).

Results: For both phantom and NHP studies, the developed SR reconstruction method yielded PET images with visibly increased spatial resolution as compared to static acquisitions reconstructed using 1 or 2 mm voxels, allowing for improved visualization of small cortical and subcortical brain phantom structures (Fig. 1 and 2). Overall, the SR method achieved better noise control than the static reconstruction with the same voxel size, owing to interpolation effects introduced by the continuous application of motion transformations during image estimation. Line profiles confirmed improved spatial resolution for the SR images in both studies. For the Hoffman phantom, the SR images showed improved correspondence with the high-resolution CT compared to the conventional methods.

Conclusions: The results demonstrate that super resolution can be achieved in brain PET by measuring object movement in real time using a high-resolution infrared tracking camera. Acknowledgments: This work was supported in part by U01EB027003 and P41EB022544. References: [1] Kennedy et. al, IEEE Trans. Med. Imag.,25(2), pp. 137-147, 2006.

• Download figure
• Open in new tab
• Download powerpoint