Variable Weighted Ordered Subset Image Reconstruction Algorithm

Pan Jin-xiao, Tie Zhou, Yan Han, Ming JiangInternational Journal of Biomedical Imaging 2006 · DOI

TL;DR

This paper introduces two novel image reconstruction algorithms, VW-ART and VW-OS-SART, that improve upon existing ART and SART methods by dynamically weighting projection data based on ray geometry. These algorithms achieve better image quality and faster convergence compared to their predecessors, as demonstrated on both simulated and real CT data.

What Problem Does This Paper Address?

Image reconstruction, particularly in fields like Computed Tomography (CT), aims to recreate a 2D or 3D image from a series of 1D projections. This process is inherently ill-posed due to noisy and incomplete data, making direct solutions infeasible. Iterative methods like Algebraic Reconstruction Technique (ART) and Simultaneous Algebraic Reconstruction Technique (SART) are commonly used, but they can be computationally expensive and may not always yield the highest quality images. Ordered Subset (OS) techniques were developed to accelerate convergence by processing subsets of the projection data. However, these methods often use fixed parameters. This paper addresses the limitations of existing iterative reconstruction algorithms by proposing a more adaptive approach that accounts for the geometrical direction of the rays, aiming to improve both image quality and convergence speed.

Methods

The core innovation lies in the introduction of "variable weighted" schemes for both ART and OS-SART. The proposed algorithms, VW-ART and VW-OS-SART, dynamically adjust the weights assigned to projection data based on the geometrical direction of the ray.

  • Subset Partitioning: The projection data is first partitioned into $T$ subsets ($B_t$) based on angle intervals. This is a standard approach in OS methods.
  • Variable Weighting: The key novelty is the introduction of weights ($\mu_{B_t}(y_i)$) that vary with the projection angle. These weights are designed to be piece-wise linear and sum to 1 for each projection datum, reflecting the geometrical direction of the ray. The paper illustrates this with a geometrical interpretation where weights are assigned based on which angle subinterval a projection falls into.
  • Algorithm Derivation: The paper derives the update equations for VW-OS-SART and VW-ART based on a general Landweber scheme. The VW-OS-SART algorithm incorporates the variable weights into the OS-SART framework, where each subset is processed iteratively. The VW-ART algorithm is presented as a variable weighted version of ART, where the relaxation parameters are adjusted based on the weighting scheme.
  • Convergence Analysis: The paper establishes convergence properties for the proposed algorithms, showing that they converge to a solution even with inconsistent data, similar to existing methods but with the added benefit of improved quality and speed.

Datasets

The paper evaluates the proposed algorithms using two types of datasets:

  • Numerical Simulation: The Shepp-Logan phantom, a standard benchmark for image reconstruction algorithms, was used. Pseudo-random Gaussian noise was added to simulate realistic noisy data.
  • Real CT Data: Experimental data from a cylindrical object with apertures was used. This dataset provides a more practical evaluation of the algorithms' performance on real-world imaging scenarios.

The paper does not introduce new benchmark datasets.

Key Insights and Ideas

  • Geometry Matters: The most significant insight is that incorporating the geometrical direction of the rays through variable weighting can substantially improve image reconstruction quality. This suggests that a more nuanced understanding of the projection process leads to better results.
  • Adaptive Weighting for Better Quality: By allowing weights to vary, the algorithms can better handle different projection angles, leading to a more accurate representation of the underlying image structure. This is particularly evident in the improved resolution of detected apertures in the practical data experiments.
  • Faster Convergence with Improved Quality: The variable weighting scheme not only enhances image quality but also accelerates the convergence of the iterative process. This is a crucial practical advantage, as it reduces the computational time required to achieve a satisfactory reconstruction.
  • Intuitive and Implementable: The proposed weighting strategy is described as intuitively reasonable and easy to implement, making it a practical advancement for existing reconstruction pipelines.

Key Results

  • Image Quality Improvement: Experimental results show that the VW-ART algorithm produces significantly better reconstructed images than ART and OS-SART, both on the Shepp-Logan phantom and real CT data. The VW-OS-SART also shows improvements over OS-SART.
  • Convergence Speed: Both VW-ART and VW-OS-SART demonstrate faster convergence speeds compared to their non-weighted counterparts (ART and SART, respectively). Figure 3 visually supports this, showing a quicker decrease in mean square error for the VW algorithms.
  • Resolution Enhancement: In the practical CT data experiments, the VW-ART algorithm was able to detect apertures that were less discernible or undetectable with other algorithms, highlighting its superior resolution capabilities.

Limitations and Future Work

  • Subset Partitioning Impact: The authors acknowledge that the choice of subset partitioning can affect image quality. While they use a geometrically based partitioning, they suggest that other partitioning strategies might be beneficial.
  • Weighting Function Exploration: The paper uses a piece-wise linear weighting function. The authors mention that other weighting functions, such as Gaussian or exponential functions, could be explored, similar to existing OS methods.
  • Pixel Set Partitioning: The authors note that partitioning the data in the projection domain is one approach. They suggest that partitioning the pixel set in the image domain is another avenue for future investigation.
  • Generalization: While the paper focuses on CT, the principles of variable weighted reconstruction could potentially be applied to other tomographic imaging modalities.

Summary updated

Abstract

We propose two variable weighted iterative reconstruction algorithms (VW-ART and VW-OS-SART) to improve the algebraic reconstruction technique (ART) and simultaneous algebraic reconstruction technique (SART) and establish their convergence. In the two algorithms, the weighting varies with the geometrical direction of the ray. Experimental results with both numerical simulation and real CT data demonstrate that the VW-ART has a significant improvement in the quality of reconstructed images over ART and OS-SART. Moreover, both VW-ART and VW-OS-SART are more promising in convergence speed than the ART and SART, respectively.
BibTeX
@article{jinxiao2006variable,
  author = {Pan Jin-xiao and Tie Zhou and Yan Han and Ming Jiang},
  title = {{Variable Weighted Ordered Subset Image Reconstruction Algorithm}},
  journal = {International Journal of Biomedical Imaging},
  year = {2006},
  doi = {10.1155/ijbi/2006/10398},
}

Loading related content...

Figure