UCSF

Cone-beam computed tomography (CBCT) has become integral to image-guided interventions and radiotherapy owing to its compact geometry and capacity for in-room volumetric imaging. Its broad-beam acquisition, however, generates substantial scattered radiation that violates the line-integral assumption underlying reconstruction, biasing attenuation estimates and degrading Hounsfield-unit (HU) accuracy, contrast-to-noise ratio (CNR), and spatial resolution. Existing scatter-correction strategies each impose characteristic penalties: hardware rejection methods such as anti-scatter grids attenuate primary fluence and raise dose; beam-stop arrays sacrifice primary signal and often require dual scans; and software approaches based on analytical kernels, Monte Carlo simulation, or machine learning trade computational cost, prior-knowledge requirements, or robustness for accuracy. This thesis introduces sparse primary sampling (SPS), a hybrid scatter estimation and correction strategy designed to recover scatter while preserving the primary signal in a single, dose-neutral acquisition. The SPS grid comprises a radiolucent holder populated with small focused tungsten collimators that expose a sparse lattice of “smart pixels” to near-primary-only signal, while the surrounding detector records the full primary-plus-scatter signal. Because Compton scatter in the diagnostic energy range is spatially smooth and low-frequency dominant, the scatter sampled at these points can be interpolated across the detector and refined through an alternating-minimization reconstruction that jointly estimates the image and the scatter distribution under edge-preserving and smoothness regularization. The method was validated computationally using GPU-accelerated Monte Carlo simulations of a cylindrical CT phantom and patient-derived head-and-neck and pelvic phantoms spanning low and high scatter-to-primary regimes, and benchmarked against a three-dimensional Richardson–Lucy denoising approach. A sampling density of approximately 0.08% of detector pixels was identified as optimal, balancing scatter sampling against interpolation error and yielding up to a 94.5% reduction in HU root-mean-square error for the CT phantom, alongside substantial cupping-artifact and CNR recovery, without dose overhead or prior anatomical knowledge. A prototype SPS grid was further evaluated experimentally on an in-house benchtop CBCT system using a Catphan phantom, comparing filtered back projection and total-variation-regularized iterative reconstruction with and without SPS correction. Despite geometric misalignment and prototype limitations that constrained absolute performance, SPS correction reduced cupping non-uniformity and improved HU linearity and CNR, with iterative reconstruction showing the strongest recovery, demonstrating proof of concept under experimental conditions. Collectively, these results aim to establish SPS as a dose-efficient, single-scan scatter-correction technique that requires no patient-specific priors and integrates measurement, estimation, and reconstruction within a unified workflow. The thesis concludes by identifying the engineering refinements and expanded phantom studies needed to close the gap between the demonstrated experimental performance and the technique’s computationally established potential.