Date of Award
Summer 2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Mathematics, Statistics and Computer Science
First Advisor
Gregory Ongie
Second Advisor
Anne Clough
Third Advisor
Emil Sidky
Fourth Advisor
Gregory Ongie
Fifth Advisor
Sarah Hamilton
Abstract
Dedicated breast X-ray computed tomography (breast CT) is a developing imaging modality that has potential as an alternative to mammography. Unlike traditional mammography, breast CT offers high-resolution, fully 3D anatomical detail, which aids in the detection of microcalfications and subtle lesions that are potential indicators of cancer. To be effective for screening, breast CT must limit radiation exposure by reducing CT view angles. However, traditional image reconstruction methods yield noisy, artifacted images in this sparse-view CT setting. This dissertation develops image restoration and reconstruction methods that address these limitations by integrating physics-informed iterative methods with machine learning tools, and a novel neural network training approach that aligns training objectives with clinically relevant signal detection tasks arising in breast CT. To recover material-specific tissue maps from nonlinear dual-energy breast CT data, an unrolled estimator that embeds a learned network as both initialization and regularizer within a fixed number of iterations of a model-based reconstruction algorithm is proposed. On simulated dual-energy breast CT data, the unrolled approach achieves superior pixel-wise image quality at lower computational cost than full iterative reconstruction. To address the misalignment between pixel-wise training losses and clinical signal detection tasks in breast CT, a novel training objective function, the signal promoter (SigPro) loss, is introduced. The SigPro loss embeds a synthetic binary signal detection task into the neural network training. Evaluated on a textured digital breast CT phantom, denoising networks trained with SigPro loss achieve higher signal detection performance than those trained with a more traditional pixel-wise loss. To enforce data consistency while preserving task-informed image priors, a pre-trained SigPro denoising network is embedded into an iterative reconstruction algorithm. The resulting estimator achieves signal detectability performance approaching the theoretical limit, substantially outperforming stand alone denoising networks and a related estimator based on a pixel-wise loss trained denoiser. To adapt the proposed image restoration methods to clinically realistic settings, a 3D fusion network is proposed and evaluated on high resolution anatomically realistic breast phantoms, providing proof-of-concept for extending task-informed deep learning based CT image restoration from 2D slices to 3D volumes