Date of Award
Summer 7-11-2023
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Over the past few decades, dual-energy X-ray computed tomography (DECT) imaging has undergone significant advancements, mainly attributed to its remarkable ability to differentiate between different materials. In contrast to conventional X-ray computed tomography (CT), DECT employs two measurements acquired at distinct spectra, enabling the generation of highly informative and precise images, including material decomposition maps, virtual non-contrast images, and stopping-power maps for proton treatment plans. Image-domain decomposition (IDD) is a widely utilized technique in DECT imaging, assuming that low- and high-energy images can be represented as linear combinations of material decomposition maps on an element-wise basis. On the other hand, DECT statistical iterative reconstruction (SIR) is an alternative DECT imaging technique that simultaneously reconstructs DECT images in both the image and sinogram domains by minimizing a penalized log-likelihood function. DECT SIR has demonstrated superior performance compared to image-domain methods, offering high accuracy and low uncertainty. Recent studies have shown that DECT SIR enables more precise stopping-power estimation with subpercentage uncertainty, outperforming the clinical practice that has an uncertainty range of 2-3.5%. In this dissertation, we present the implementation of a GPU-accelerated 3D DECT SIR for helical measurements and demonstrate that its performance is comparable to the CPU-based 2D DECT SIR for axial data. The simulation results indicate that both the 2D CPU-based and 3D GPU-based implementations yield results within a margin of error of 0.2% for all inserts. Thanks to the accelerated GPU implementation, the total elapsed time of DECT SIR has been reduced by a factor of 1577 relative to the single-threaded CPU implementation. However, during the reconstruction of clinical measurements, several issues related to clinical scenarios have been observed. These include motion artifacts resulting from patient movement during sequential DECT scans, artifacts caused by missing or corrupted measurements, and clinically unacceptable time-to-solution even if multiple acceleration techniques have been exploited. Sequential scanning is a popular dual-energy data acquisition method as it requires no specialized hardware. However, patient motion between two sequential scans may lead to severe motion artifacts in DECT SIR images. In order to reduce the motion artifacts in such reconstructions, we propose a motion-compensation scheme that incorporates the deformation vector field into any DECT SIR. The deformation vector field is estimated via the multi-modality symmetric deformable registration method. The precalculated registration mapping and its inverse or adjoint are then embedded into each iteration of the iterative DECT algorithm. A perturbation analysis was then performed to determine errors in approximating the continuous deformation by using the deformation field and interpolation. Results from a simulated and clinical case show that the proposed framework is capable of reducing motion artifacts in DECT SIR. Percentage mean squared errors in regions of interest in the simulated and clinical cases were reduced from 4.6% to 0.5% and 6.8% to 0.8%, respectively. Furthermore, incomplete and corrupted measurements commonly occur in clinical settings due to the presence of metal implants and the small field-of-view (FOV) of the scanner. These incomplete or corrupted measurements introduce severe streaking artifacts during DECT SIR, which can compromise diagnostic accuracy and treatment performance. To mitigate such artifacts, we introduce a scheme for reconstructing incomplete or corrupted data. During reconstruction, the corrupted or missing parts are ignored and substituted by the mean polychromatic transmission. In scenarios involving metal artifact reduction (MAR), we utilize a normalized metal-artifact reduction (NMAR) technique that combines image-domain decomposition to initialize the algorithm and expedite convergence. Comparative visualization and quantitative analysis reveal that DECT SIR with the proposed method outperforms other techniques in reducing streaking artifacts caused by metallic objects. In an experimental case, the mean absolute errors are reduced from 6% to within 1% in the energy range of 60 to 150 keV. Moreover, the performance of DECT SIR for incomplete data is evaluated using clinically acquired measurements with simulated small FOV. With the FOV extension technique, the mean absolute errors are reduced from 1.87% and 1.33% to 0.73% and 0.58% at 60 and 150 keV, respectively. The slow convergence rate of DECT SIR and the substantial computational demands of projections often result in clinically unacceptable time-to-solution. To address this, we have integrated DECT SIR into a deep-learning model-based unrolled network for 3D DECT reconstruction (MB-DECTNet), which can be trained end-to-end. This deep-learning-based approach is designed to learn shortcuts between initial conditions and the stationary points of iterative algorithms, while preserving the unbiased estimation property of model-based algorithms. MB-DECTNet consists of multiple stacked update blocks, each comprising a data-consistency layer (DC) and a spatial mixer layer. The DC layer serves as a one-step update for any traditional iterative algorithm. The quantitative results indicate that our proposed MB-DECTNet surpasses both the traditional image-domain technique (the average bias is decreased by a factor of 10) and a pure deep-learning method (the average bias is decreased by a factor of 8.8), offering the potential for accurate attenuation coefficient estimation, akin to traditional statistical algorithms, but with considerably reduced computational costs. This approach achieves 0.13% bias and 1.92% mean absolute error and reconstructs a full image of a head in less than 6 minutes. Additionally, we show that the MB-DECTNet output can serve as an initializer for DECT SIR, leading to further improvements in results.
Language
English (en)
Chair
Joseph O’Sullivan