Date of Award
Fall 2019
Document Type
Dissertation - Restricted
Degree Name
Doctor of Philosophy (PhD)
Department
Biomedical Engineering
First Advisor
Schmidt, Taly G.
Second Advisor
Clough, Anne
Third Advisor
Sidky, Emil
Abstract
Energy-resolved photon counting detectors have potential improvements over conventional scintillating x-ray detectors and can be used for spectral computed tomography (CT) applications such as material decomposition. However, photon counting detector are difficult to model because they have many non-ideal physical phenomena. We developed machine learning algorithms to model the spectral measurements of energy-resolved photon counting detectors for the purposes of material decomposition and image reconstruction. A feed-forward neural network was used to estimate path lengths of different materials that the x-ray path traversed from spectral measurements of a photon counting detector using calibration measurements on an experimental bench-top CT system. The neural network relied only on calibration measurements and no prior knowledge of the source spectrum, material attenuation coefficients, or detector response. The neural networks decomposed each ray in the measured sinograms into material separated sinograms and reconstructed into basis material images. The capabilities of the machine learning algorithm were evaluated experimentally with a rod phantom with calculated biases of 0.3%to 7.6% compared to biases of 1.3% to 16% of a previously proposed empirical decomposition algorithm. A special case of material decomposition occurs when materials with high atomic numbers are in the measured ray path. These materials have K-absorption edges in their linear attenuation coefficients in the relevant x-ray energy range. In a separate study, the neural network mode was extended for K-edge imaging to separate contrast agents and nanoparticles, such as iodine and gadolinium, from other materials in the image. Transfer learning was used to further train the neural networks between detector pixels and from simulated calibration data to improve the qualitative and quantitative image quality in a rod phantom and an ex-vivo rat leg specimen. Similarly, a neural network was developed to model the forward spectral measurement from basis material parameters of the object. The forward model was able to fit the calibration measurements to within 6% of all measurement and was used with an optimization algorithm for material decomposition. Such a model could be used for iterative reconstruction algorithms and iterative material decomposition algorithms to optimize the estimates using the noise properties of the measurement.