Strategies for Practical, High Quality and Accurate List Mode Iterative Image Reconstruction for PET

The 1 mm resolution clinical PET system under development utilizes 1x1x1 mm3 crystal elements integrated into two panels, each measuring roughly 10x16 cm2, that form several billion possible annihilation photon response lines (LORs) between pairs of individual elements from each panel. To avoid a potentially enormous memory burden we chose to perform the required line projection operations, the critical and most computational steps of iterative tomographic image reconstruction, using on-the-fly calculations, on a per event basis. To address this problem we drew from advances in the computer graphics field. It turns out that these line projection operations can be reformulated as geometric calculations that can run much more efficiently on graphics hardware known as the graphics processing unit (GPU) rather than running software on a cluster of CPUs. As a result of this novel reformulation of the problem, the iterative image reconstruction algorithm, known as 3-D list-mode ordered subsets expectation maximization (3D-OSEM), runs a factor of 50 to 100 times faster on a GPU than an equivalent-cost state-of-the-art CPU, thus replacing the need for a large CPU cluster, and reducing associated costs and physical footprint, in order to make the system more available and practical for use in the tight confines of the breast imaging clinic, while maintaining equivalent or better image quality/quantitative accuracy.

In addition to high resolution PET, the techniques are being developed to handle list-mode 3D-OSEM image reconstruction for time-of-flight (ToF) PET.

People Involved