Task-informed Metrologies for Characterization and Optimization in Spectral CT

Abstract

The purpose of this dissertation was to develop methods for characterization and optimization of spectral CT for clinical usage. This was done by focusing on the information content in spectral CT image datasets and how that information is understood within each image and between images. This purpose was carried out in three stages. First, conventional task-based metrologies were applied to photon-counting CT to assess the improvements offered by the technological advancement over energy-integrating detector CT systems for a series of specific clinical applications. Second, task-based metrologies were developed for spectral CT by focusing on the multi-dimensional nature of signal and information in multi-energy imaging. Third, the utility of the developed metrologies was demonstrated by application to questions of material characterization and decomposition.In part 1, a prototype photon-counting CT system was compared to clinical energy integrating CT for a range of clinical tasks. Using a multi-tiered image quality phantom, both systems were compared for performance for soft tissue, vascular, and high-resolution imaging. In another study, an image quality phantom and vials of iodine contrast were imaged with both systems to characterize reduced dose performance. Finally, a coronary plaque and stent phantom was imaged with both systems and evaluated with qualitative and quantitative methods for cardiac imaging tasks. Across the requirements for each clinical tasks, the benefits of photon-counting CT suggested comparable or superior performance compared to conventional energy integrating CT. In part 2, two metrologies were developed by incorporating the multi-dimensional nature of information in spectral CT. The first metrology was an extension of the signal-to-noise ratio, a task generic conventional metric, into the multivariate domain. The multivariate signal-to-noise ratio characterized image quality within energy channels by including the conventional term and characterized the influence of channels on one another with a covariance weighted signal-to-noise ratio term. The second metrology was a method to measure the overlap between two signals of interest which were defined over multiple energy levels. The separability index provided a task specific method to characterize the separability of signals in spectral CT image datasets. Both metrologies provided additional characterization of spectral CT performance that was not possible with conventional image quality metrics. In part 3, the two metrologies were applied to answer questions regarding material characterization and material decomposition within spectral CT using simulated data. A series of materials, including candidate contrast agents and nanomaterials, were imaged across a range of energy threshold measures. The relationship between signal and separability with energy threshold was characterized between materials. Another dataset containing iodine and gadolinium was generated across a range of image formation parameters. Both metrologies were applied to this dataset and compared to material decomposition performance across the dataspace. Finally, a deep learning approach to material decomposition was trained and validated using simulated data. The algorithm showed excellent performance for quantification of iodine and gadolinium. Both metrologies were shown to be useful to characterize the relationship between image formation factors and material signal properties and how those considerations can shape development of material decomposition algorithms. In conclusion, this dissertation provides task-informed metrologies that can be used for characterization, evaluation, and optimization of spectral CT systems for signal dependent processing tasks including material identification and quantification.