A Highly Accelerated Parallel Multi-GPU based Reconstruction Algorithm for Generating Accurate Relative Stopping Powers

TL;DR

This paper introduces a novel parallel multi-GPU reconstruction algorithm for low-dose Proton Computed Tomography that achieves high accuracy and speed on single GPU systems, reducing reliance on large clusters.

Contribution

The paper presents a new GPU-based reconstruction technique for pCT that is highly accelerated and effective on single GPU hardware, unlike previous methods requiring extensive clusters.

Findings

Achieves accurate RSP maps within clinically acceptable time frames.

Operates efficiently on single GPU systems like Nvidia P100.

Demonstrates effectiveness on both simulated and real datasets.

Abstract

Low-dose Proton Computed Tomography (pCT) is an evolving imaging modality that is used in proton therapy planning which addresses the range uncertainty problem. The goal of pCT is generating a 3D map of Relative Stopping Power (RSP) measurements with high accuracy within clinically required time frames. Generating accurate RSP values within the shortest amount of time is considered a key goal when developing a pCT software. The existing pCT softwares have successfully met this time frame and even succeeded this time goal, but requiring clusters with hundreds of processors. This paper describes a novel reconstruction technique using two Graphics Processing Unit (GPU) cores, such as is available on a single Nvidia P100. The proposed reconstruction technique is tested on both simulated and experimental datasets and on two different systems namely Nvidia K40 and P100 GPUs from IBM and Cray. The experimental results demonstrate that our proposed reconstruction method meets both the timing and accuracy with the benefit of having reasonable cost, and efficient use of power.