Scatter Estimation and Correction for Experimental and Simulated Data in Multi-Slice Computed Tomography Using Machine Learning and Minimum Least Squares Methods

Author

Cornelia WangFollow

ORCID

https://orcid.org/0000-0002-2692-3538

Date of Award

Summer 8-19-2021

Author's School

McKelvey School of Engineering

Author's Department

Electrical & Systems Engineering

Degree Name

Master of Science (MS)

Degree Type

Thesis

Abstract

Current research aims to reduce the stopping power ratio prediction error in the inputs to the proton therapy planning process to less than 1%, which allows for improved radiation therapy planning. Our present study on reducing SPR error neglects the effect of scattering, which can increase SPR error by as much as 1-1.5%. The idea is that for each source-to-detector pair, 24 mm collimation data is close to 3 mm collimation data but with increased signal due to scattering. The goal is to estimate 3 mm collimation data from 24 mm collimation data. Pairs of sinograms, both experimental data and simulated data, from 3 mm and 24 mm collimation data are used to derive methods for this scatter correction. One method uses a linear least-squares approach to derive a linear estimator. A second method uses a U-net structure in a machine learning approach. An experiment is run using Monte Carlo simulation data to predict the 3 mm scatter-only signal from the 24 mm scatter-only signal using least squares estimation. The current version of the U-net structure cannot predict scatter-corrected data successfully because more artifacts are introduced. The proposed least squares model can use local measurements to estimate scatter locally. In 2 of 3 groups of phantom data, reconstructed images of scatter- corrected data show higher uniformity and structural similarity with the ground truth than uncorrected data. The highest Structural Similarity Index Measure reaches 0.9869, and the lowest nonuniformity index reaches 2.16%. My study found that using local measurements to estimate scatter locally, the least-squares model keeps corrected sinogram low error and significantly improves the quality of corrected images by observing fewer artifacts, lower nonuniformity index, and higher structural similarities.

Language

English (en)

Chair

Joseph A. O'Sullivan

Committee Members

Jeffrey F. Williamson David G. Politte Joseph P. Culver