Program Information
A Fully-Automated, High-Throughput, Reconstruction and Analysis Pipeline for Quantitative Imaging in CT
J Hoffman*, M Wahi-Anwar , N Emaminejad , H Kim , M Brown , M McNitt-Gray , UCLA School of Medicine, Los Angeles, CA
Presentations
TU-C2-GePD-IT-2 (Tuesday, August 1, 2017) 10:00 AM - 10:30 AM Room: Imaging ePoster Theater
Purpose: Cohort size and number of reconstructions often limit the statistical power or clinical applicability of quantitative imaging (QI) studies. This is often due to the logistical overhead of case collection, performing large numbers of reconstructions on-scanner, and management of hundreds or thousands of reconstructed datasets. In this work, a high-throughput, fully-automated CT image reconstruction and analysis “pipeline” is described that overcomes the limitations of previous approaches and advances QI analysis capabilities.
Methods: The pipeline can be divided into two primary components: reconstruction and analysis. The reconstruction portion takes CT projection data, applies any raw-data preprocessing such as filtration or noise addition, and performs reconstruction. Output is the reconstructed image data and metadata. The analysis portion accepts reconstructed images and performs tasks, such as segmentation, computer-automated-detection (CAD), emphysema scoring, or texture feature analysis. Full-automation is achieved via supervisor scripts written in Python, and high-throughput is achieved using GPU-acceleration and cluster computing. Currently, FreeCT_wFBP, an implementation of weighted-filtered-backprojection, is the primary reconstruction technique, although an implementation of fully-3D model-based-iterative reconstruction is under development.
Results: The pipeline was initially tested with a pilot study into the robustness of emphysema scoring to CT parameter selection using 360 reconstructions (10 patients, 36 parameter configurations per patient). Total time required to perform reconstructions was 13.8 hours and analysis required approximately 40 minutes. A similar pilot study focusing on CAD utilized the pipeline to generate and analyze 1800 reconstructions (50 patients, 36 reconstructions per patient). Reconstruction required 2.7 days (65.15 hours) and CAD execution required ~20 hours. Supervisor scripts effectively reduced computational latency to zero, and eliminated previously-required human intervention after initial configuration.
Conclusion: The full-automation and high-throughput nature of the pipeline represents a paradigm shift in QI research and has enabled new types of studies and accelerated existing study workflows significantly.
Funding Support, Disclosures, and Conflict of Interest: Disclosures: *All authors - Institutional research agreement, Siemens Healthineers Funding support: *University of California Office of the President Tobacco-Related Disease Research Program (UCOP-TRDRP grant #22RT-0131) *National Cancer Institutes Quantitative Imaging Network (QIN grant U01-CA181156)
Contact Email: