A Preconditioned Algorithm for Model-Based Iterative CT Reconstruction and Material Decomposition from Spectral CT Data

Matthew Tivnan, Wenying Wang, J. Webster StaymanarXiv 2020 · DOI · arXiv

Abstract

Model-based material decomposition is a statistical iterative reconstruction framework where basis material density images are estimated directly from spectral CT data. This method uses a physical model for polyenergetic x-ray transmission and attenuation and therefore it does not typically suffer from beam-hardening artifacts. However, this estimation is a poorly-conditioned inverse problem due to the strong anticorrelation between basis materials. In this work we propose a preconditioned optimization algorithm for a nonlinear penalized weighted least-squares objective function.

Introduction

Model-Based Iterative Reconstruction (MBIR) represents the current gold standard for image quality in Computed Tomography (CT). By incorporating sophisticated physical models of X-ray physics—such as beam-hardening and noise statistics—MBIR enables significant radiation dose reduction while maintaining high diagnostic quality. However, the computational burden of MBIR remains a primary barrier to its widespread clinical adoption, particularly in the context of Spectral CT.

In Spectral CT, the challenge is amplified by the need for "material decomposition," the process of separating an image into constituent components like bone, iodine, or soft tissue. Mathematically, this is a "poorly conditioned" inverse problem. Because different materials often exhibit similar X-ray attenuation profiles, optimization algorithms struggle to distinguish between them. This results in a strong "anticorrelation" between material maps, leading to noisy estimates and requiring thousands of iterations to converge.

The Preconditioning Framework

This research proposes a preconditioned version of the Separable Parabolic Surrogates (SPS) algorithm to address these bottlenecks. While SPS is favored for its guaranteed monotonic convergence, standard implementations often rely on simple diagonal scaling. The authors argue that diagonal preconditioners are too conservative and fail to capture the complex geometry of the optimization landscape in Spectral CT.

To overcome this, the paper derives novel linear transformations (preconditioners) that redefine the estimation space. The core contribution lies in the derivation of these preconditioners for four distinct scenarios:

  1. Standard CT: Utilizing a shift-invariant Fourier-based preconditioner.
  2. Shift-Variant CT: Implementing a Haar-wavelet-based preconditioner to handle complex geometries like cone-beam CT.
  3. Projection-Domain Decomposition: A cross-material block-diagonal preconditioner.
  4. One-Step MBMD: A sophisticated cross-material, cross-voxel preconditioner using either Fourier or Wavelet bases.

Key Methodological Insights

Fourier vs. Wavelet Preconditioning

The authors utilize the Discrete Fourier Transform (DFT) to approximate the Hessian (the matrix of second derivatives) of the objective function. This approach is highly effective for shift-invariant systems where the response is uniform across the field of view. However, for shift-variant systems, the paper introduces Haar wavelets. Unlike Fourier transforms, wavelets encode information about both frequency and spatial location, making them better suited for the complex geometries of modern scanners.

Resolving Material Anticorrelation

A significant innovation in this work is the "Cross-Material Preconditioning" strategy. This accounts for the fact that an update in one material image (e.g., bone) should mathematically inform the update in another (e.g., soft tissue). By smoothing the optimization landscape across material dimensions, the algorithm can resolve inherent anticorrelations in far fewer steps than traditional methods.

Discussion and Future Work

While the paper is primarily a theoretical derivation, it provides the necessary mathematical framework to make MBIR practical for real-time clinical use. The authors identify a critical trade-off: while preconditioning reduces the total number of iterations required for convergence, each individual iteration becomes more computationally expensive due to the need for transforms like the Fast Fourier Transform (FFT).

Future work will involve the empirical validation of these algorithms on high-resolution clinical spectral CT data. This will help quantify the exact speedup and determine the optimal balance between wavelet resolution and computational overhead. By providing these mathematical "levers," the authors move the field closer to a version of MBIR that is both high-quality and computationally efficient.

Summary updated

Abstract

Model-based material decomposition is a statisticaliterative reconstruction framework where basis material densityimages are estimated directly from spectral CT data. This methoduses a physical model for polyenergetic x-ray transmission andattenuation and therefore it does not typically suffer frombeam-hardening artifacts. However, this estimation is a poorly-conditioned inverse problem due to the strong anticorrelationbetween basis materials. In this work we propose an precondi-tioned optimization algorithm for a nonlinear penalized weightedleast-squares objective function.
BibTeX
@misc{tivnan2020preconditioned,
  author = {Matthew Tivnan and Wenying Wang and J. Webster Stayman},
  title = {{A Preconditioned Algorithm for Model-Based Iterative CT Reconstruction and Material Decomposition from Spectral CT Data}},
  year = {2020},
  eprint = {2010.01371},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.01371},
}

Loading related content...

Figure