GPU implementation of list-mode DRAMA for real-time OpenPET image reconstruction

Abstract

The OpenPET, which has a physically opened space between two detector rings, is our new geometry to enable PET imaging during radiation therapy. Especially, tracking a moving target such as a tumor in the lung will become possible if the real-time imaging system is realized. In this paper, we developed a list-mode image reconstruction method using general-purpose computing on graphics processing units (GPGPUs) toward real-time image reconstruction. We used the dynamic row-action maximum likelihood algorithm (DRAMA). For GPU implementation, efficiency of acceleration depends on implementation methods; reduced conditional statements and efficient memory accesses are required. On the other hand, accurate system model is required to improve quality of reconstructed images. Therefore, we developed a new system model which was suited for the GPU implementation. In the new system model, the detector response functions (DRFs), which were calculated analytically to represent the probability distribution of each LOR, were modeled by sixth-order polynomial functions. The system model enabled us to calculate each element of the system matrix with reduced conditional statements. We applied the developed method to a small OpenPET prototype, which was developed for a proof-of-concept. We compared the proposed system model to the sub-LOR model, a geometrically-defined accurate system model which we had previously proposed. The difference between the reconstructed images with the new system model using GPU and the sub-LOR model using CPU was very small. Our new system model on GPU was 46.3 times faster than the sub-LOR model on CPU.