Author: Ehsan Abadi, Zakaria Aboulbanine, Nicholas D Felice, David Fenwick, Anuj J. Kapadia, Cindy Marie McCabe, Jayasai Ram Rajagopal, Ehsan Samei 👨🔬
Affiliation: Duke University, Oak Ridge National Laboratory, Center for Virtual Imaging Trials, Duke University, Clinical Imaging Physics Group, Department of Radiology, Duke University Health System 🌍
Purpose:
Virtual imaging trials (VITs) offer a computational alternative to clinical imaging trials leveraging virtual patients, scanners, and interpreters to assess imaging questions. To provide meaningful insights, VITs must incorporate large datasets capturing the clinical population variability and imaging acquisition parameters. Conducting large-scale trials necessitates the use of high-performance computing (HPC) resources for efficiency and scalability. In this study, we implemented a large-scale computed tomography (CT) VIT utilizing the supercomputing capabilities at Oak Ridge National Laboratory (ORNL).
Methods:
The study population included virtual phantoms (XCATs) representing different patient populations including adult and pediatric patients with lung lesions, adult patients with chronic obstructive pulmonary disease (COPD), and adult patients with liver lesions. Patient models were imaged using a validated simulation platform (DukeSim) that combines CPU and GPU-driven processes to estimate primary and scatter signals. Scan protocols for each population were adapted to standard clinical ranges for each patient population across tube voltage and tube current conditions. Simulations were performed using Summit, a state-of-the-art supercomputer featuring 4600 compute nodes, each equipped with two IBM POWER9 processors and six NVIDIA Tesla V100 GPUs.
Results:
Across all datasets and acquisition conditions, a total of 10,140 simulations were conducted, generating over 125 TB of CT projection data. Each simulation was executed on individual compute nodes, with runtimes between 35-100 minutes and an average run time of 45 minutes per simulation. On a single node, a VIT of this scale would take 317 days to complete. Leveraging the HPC resources at ORNL, the study was completed within a two-week period – an unprecedented efficiency compared to previous simulation studies or clinical trials of smaller scale.
Conclusion:
By leveraging HPC resources, we successfully conducted a population scale VIT, generating a dataset of over 10,000 simulations to address large-scale clinical questions.