Optimizing Parameters for Enhanced Iterative Image Reconstruction Using Extended Power Divergence

Takeshi Kojima, Yusaku Yamaguchi, Omar M. Abou Al-Ola, Tetsuya YoshinagaAlgorithms 2024 · DOI

TL;DR

This paper introduces a novel method to automatically optimize parameters in iterative image reconstruction algorithms, specifically the Power Divergence-based Expectation-Maximization (PDEM) method. By optimizing an evaluation function based on extended power divergence measures, the approach achieves high-quality image reconstruction comparable to using the true image, even with noisy projection data, without relying on machine learning.

What Problem Does This Paper Address?

Medical imaging techniques like X-ray computed tomography (CT) and nuclear medicine CT rely on reconstructing 2D or 3D images from projection data. While transform methods are fast, they struggle with noisy or incomplete data, leading to artifacts. Iterative reconstruction methods, such as Maximum Likelihood Expectation-Maximization (MLEM), offer superior image quality and noise robustness, making them crucial for reducing radiation doses. However, these iterative methods often have adjustable parameters that significantly impact reconstruction performance. Finding the optimal parameter values is challenging, with existing methods like machine learning or sequential estimation having limitations. This paper addresses the need for an efficient and effective way to determine these optimal parameters for iterative reconstruction algorithms, particularly for the PDEM method, to achieve high-quality images under realistic imaging conditions.

Methods

The core of this paper is a proposed method for optimizing the parameters of iterative reconstruction algorithms. The authors focus on the PDEM algorithm, which uses an extended power divergence (EPD) measure as its objective function and has two adjustable parameters: $\gamma$ and $\alpha$.

The proposed method works as follows:

  1. Define an iterative reconstruction system: For a set of $K$ projection datasets ($Y_k$) with common characteristics (e.g., similar noise levels), an iterative algorithm $f$ is defined to reconstruct pixel values $Z_k(n)$ at iteration $n$. The algorithm's behavior is governed by a parameter vector $\lambda = (\gamma, \alpha)$.
  2. Iterative Reconstruction: Starting from an initial guess $Z_k(0)$, the algorithm iterates $N$ times to produce $Z_k(N)$ for each projection dataset.
  3. Optimization Objective: An evaluation function $g(\lambda; N)$ is defined as the average of a chosen distance function $d$ applied to the reconstructed images $Z_k(N)$ and their corresponding projections. The goal is to find the parameter set $\lambda$ that minimizes $g(\lambda; N)$.
  4. Distance Functions: The paper explores several distance functions:
  5. Image-based: Euclidean distance between the reconstructed image and the true image ($d(x) = ||e - x||$).
  6. Projection-based: Euclidean distance, KL-divergence, Extended Power Divergence (EPD), and Weighted Extended Power Divergence (WEPD) between the measured projection $Y_k$ and the forward projection $AZ_k(N)$.

The optimization is performed using MATLAB's fmincon solver, with initial parameter values set to $\lambda(0) = (1, 1)$ (which corresponds to the MLEM method).

Datasets

The paper utilizes both simulated and physical datasets for evaluation:

  • Simulated Datasets:
    • Disc Phantom: A simple geometric phantom.
  • Modified Shepp-Logan Phantom: A standard phantom used in medical imaging. These phantoms were used to generate simulated projection data with controlled noise levels (signal-to-noise ratio of 20 dB) to mimic quantum and electronic noise in X-ray CT.

  • Physical Dataset:

    • Abdominal Phantom: A phantom simulating human body density variations, acquired using a Canon Medical Systems X-ray CT scanner. This dataset was chosen to investigate the noise suppression capabilities of the PDEM method under relatively high measurement noise conditions. Projections were subsampled to simulate sparse data.

Key Insights and Ideas

  • Parameter Optimization is Crucial: The study strongly emphasizes that the performance of iterative reconstruction algorithms like PDEM is highly sensitive to parameter choices.
  • Projection-Based Optimization is Viable: A key finding is that optimizing parameters using distance functions between forward projections and measured projections can yield results comparable to using the true image as a reference, which is often unavailable in real-world scenarios.
  • EPD and WEPD Shine: The Extended Power Divergence (EPD) and Weighted Extended Power Divergence (WEPD) measures, when used as distance functions for projection-based optimization, consistently lead to superior reconstruction quality compared to traditional Euclidean distance or KL-divergence.
  • Robustness to Noise: The optimized PDEM method, particularly with EPD/WEPD distance functions, demonstrates remarkable robustness against noise, achieving high-quality reconstructions even with limited projection data.
  • No Machine Learning Needed: The proposed optimization framework bypasses the need for complex machine learning models for parameter selection, offering a more direct and interpretable approach.

Key Results

  • Numerical Experiments (Disc and Shepp-Logan Phantoms):
    • When using the Euclidean distance between the reconstructed image and the true image, the optimized PDEM parameters ($\lambda(T)$) significantly reduced the evaluation function compared to the initial parameters ($\lambda(0)$).
  • Images reconstructed with optimized PDEM parameters showed less variability and sharper features compared to MLEM.
  • When using projection-based distance functions (EPD and WEPD), the optimized PDEM method achieved significantly higher SSIM scores and lower standard deviations (indicating better image quality and less noise) than MLEM and PDEM with default parameters. For the disc phantom, SSIM values reached up to 0.983 with EPD/WEPD.
  • The optimized parameters for EPD/WEPD were notably different from the initial $(1,1)$ values, often featuring a small $\gamma$ and a large $\alpha$.
  • Physical Experiment (X-ray CT Scanner):
    • The results from the physical experiment mirrored the numerical findings. The PDEM method with optimized parameters, especially using WEPD as the distance function, produced visually superior reconstructions with reduced noise and improved detail compared to MLEM.
  • Density profiles from the physical experiment confirmed the improved performance of PDEM with optimized parameters.

Limitations and Future Work

  • Dataset Dependency: The authors acknowledge that the effectiveness of the proposed method and the optimal parameter values might be dataset-dependent. While the study focused on conditions with high noise and sparse projections (motivated by dose reduction), performance might vary under low-noise conditions.
  • Computational Cost: While the method aims to optimize parameters efficiently, iterative reconstruction itself can be computationally intensive. The paper mentions using parallel computing for numerical calculations.
  • Future Work: The paper suggests that further investigation into the behavior of the proposed method under low-noise conditions is warranted. Exploring other iterative reconstruction algorithms and different distance functions could also be fruitful avenues for future research. The potential for integrating this optimization approach with other advanced techniques, like deep learning, could also be explored.

Summary updated

BibTeX
@article{kojima2024optimizing,
  author = {Takeshi Kojima and Yusaku Yamaguchi and Omar M. Abou Al-Ola and Tetsuya Yoshinaga},
  title = {{Optimizing Parameters for Enhanced Iterative Image Reconstruction Using Extended Power Divergence}},
  journal = {Algorithms},
  year = {2024},
  doi = {10.3390/a17110512},
}
Figure