SPULTRA: Low-Dose CT Image Reconstruction With Joint Statistical and Learned Image Models

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Low-dose computed tomography (CT) image reconstruction is crucial for reducing radiation exposure.
• Traditional penalized weighted-least squares (PWLS) methods, using post-log data, suffer from degraded quality in low-dose scans due to model limitations.

Purpose of the Study:

• To develop a novel CT image reconstruction algorithm that enhances image quality in low-dose scans.
• To address the limitations of post-log statistical models in PWLS by utilizing pre-log measurements.

Main Methods:

• A shifted-Poisson (SP) model-based likelihood function was investigated, utilizing pre-log raw measurements for improved statistical representation.
• A data-driven regularizer, the Union of Learned TRAnsforms (ULTRA), was incorporated, forming the SPULTRA framework.
• The SPULTRA algorithm employs quadratic surrogate functions for the SP data-fidelity term, involving iterative quadratic subproblems and sparse coding/clustering.

Main Results:

• The proposed SPULTRA algorithm demonstrated superior image reconstruction quality compared to the conventional PWLS-ULTRA method, particularly in low-dose CT scenarios.
• The computational cost per iteration of SPULTRA is comparable to PWLS-ULTRA.
• Nonconvexity in both the data-fidelity term and regularizer was addressed within the SPULTRA framework.

Conclusions:

• The SPULTRA algorithm, leveraging a shifted-Poisson model and learned transforms, offers significant improvements in low-dose CT image reconstruction.
• Utilizing pre-log measurements provides a more accurate statistical model, leading to enhanced image fidelity.
• SPULTRA presents a promising advancement for medical imaging applications requiring high-quality reconstructions from low-dose CT data.