Regularization Solver Guided FISTA for Electrical Impedance Tomography

Qian Wang, Xiaoyan Chen, Di Wang, Zichen Wang, Xinyu Zhang, Na Xie, Lili Liu2023 · DOI

TL;DR

This paper introduces a novel Regularization Solver Guided Fast Iterative Shrinkage-Thresholding Algorithm (RS-FISTA) for Electrical Impedance Tomography (EIT). By adaptively determining an initial conductivity guess using a regularization approach and incorporating Nesterov acceleration, RS-FISTA significantly improves image reconstruction quality, reduces artifacts, and achieves faster convergence compared to existing methods.

What Problem Does This Paper Address?

Electrical Impedance Tomography (EIT) is a valuable non-destructive imaging technique used in various fields, from biomedical applications like lung imaging to industrial process monitoring. However, reconstructing accurate conductivity distributions from EIT data is challenging due to the inherent nonlinearity and ill-posed nature of the inverse problem. This often leads to low spatial resolution, blurred boundaries, and artifacts in the reconstructed images. Existing iterative reconstruction methods, while effective, suffer from drawbacks such as subjective initial parameter settings, lengthy computation times, and a tendency to produce blurry images. This paper aims to address these limitations by developing a more robust and efficient EIT reconstruction algorithm.

Methods

The core of this work is the proposed Regularization Solver Guided Fast Iterative Shrinkage-Thresholding Algorithm (RS-FISTA). This method builds upon the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) by introducing two key innovations:

  • Regularization Solver Guided Initial Guess: Instead of relying on subjective experience for setting the initial conductivity matrix in FISTA, RS-FISTA employs an adaptive regularization parameter method. This approach uses a Tikhonov regularization framework with an adaptive regularization coefficient that is adjusted iteratively. This results in an initial guess that incorporates more prior information about the conductivity distribution, leading to improved accuracy and reduced artifacts.
  • Nesterov Acceleration: The algorithm incorporates Nesterov's acceleration technique into the FISTA framework. This accelerates the gradient optimization process by using a linear combination of previous iterates, leading to faster convergence and fewer iterations required to reach a satisfactory solution.

The paper also details the mathematical formulation of the EIT forward and inverse problems, including the use of Maxwell's equations and the derivation of the inverse problem model as a regularized least-squares problem.

Datasets

The paper utilizes both simulated and experimental data for evaluation.

  • Simulated Data: The authors generated simulated EIT data based on various models with different numbers and sizes of inclusions. These simulations were performed using MATLAB and COMSOL Multiphysics.
  • Experimental Data: A physical EIT tank experiment was conducted using a TUST-EIT system. The experiment involved a tank with 16 electrodes, a homogenous background, and objects with varying diameters.

The paper does not introduce new benchmark datasets but rather uses these datasets to compare the performance of RS-FISTA against established methods.

Key Insights and Ideas

  • Adaptive Regularization for Initial Guess is Crucial: The paper strongly suggests that a well-informed initial guess, derived from adaptive regularization, is more critical for EIT reconstruction quality than previously emphasized. This approach effectively injects prior knowledge into the iterative process.
  • Nesterov Acceleration Significantly Boosts Convergence: The integration of Nesterov acceleration demonstrably speeds up the convergence of the iterative reconstruction process, reducing the number of iterations needed for high-quality results.
  • RS-FISTA Outperforms Standard Methods Across Metrics: The proposed method consistently shows superior performance in terms of image quality (SSIM, PSNR) and accuracy (RMSE) across various simulation and experimental scenarios, particularly in reconstructing smaller inclusions and sharper boundaries.
  • Robustness to Noise: The RS-FISTA method exhibits good robustness to Gaussian white noise, maintaining reasonable visualization quality even at lower signal-to-noise ratios.

Key Results

The RS-FISTA method demonstrated significant improvements over other compared algorithms (Landweber, CG, NOSER, Newton-Raphson, ISTA, and FISTA) in both simulation and experimental results.

  • Simulation Experiments:
    • RS-FISTA achieved the highest SSIM (e.g., 0.7682 on average) and PSNR (e.g., 36.6065 on average), indicating better image quality and structure preservation.
  • It achieved the lowest RMSE (e.g., 3.4413 on average), signifying higher accuracy in conductivity distribution.
  • The paper reports SSIM improvements of up to 38.08% over ISTA and 33.37% over FISTA. RMSE improvements were up to 40.28% over FISTA.
  • Tank Experiments:
    • Similar to simulations, RS-FISTA yielded superior SSIM (e.g., 0.7253 on average) and PSNR (e.g., 37.5507 on average), with the lowest RMSE (e.g., 3.4413 on average).
  • SSIM improvements reached up to 38.16% over FISTA, and RMSE improvements were up to 46.05% over ISTA.
  • Convergence: Figure 6 shows that RS-FISTA achieves a stable SSIM performance within approximately 30 iterations, demonstrating faster convergence than most other iterative methods.

Limitations and Future Work

  • Limitations:
    • The paper mentions that the initial guess for the FISTA method is usually set according to experience, which has great uncertainty. While RS-FISTA addresses this with an adaptive regularization approach, the effectiveness of this adaptive regularization might still be influenced by the choice of initial regularization parameters or the specific regularization model used.
  • The paper focuses on 2D EIT models. While the principles are likely extendable, 3D EIT presents additional computational challenges.
  • Future Work:
    • The authors suggest that the RS-FISTA method can be further improved by exploring more advanced regularization techniques and acceleration strategies.
  • Extending the method to 3D EIT reconstruction is a natural next step.
  • Investigating the application of RS-FISTA to real-world EIT scenarios with complex geometries and more challenging conductivity distributions would be valuable.

Summary updated

BibTeX
@misc{wang2023regularization,
  author = {Qian Wang and Xiaoyan Chen and Di Wang and Zichen Wang and Xinyu Zhang and Na Xie and Lili Liu},
  title = {{Regularization Solver Guided FISTA for Electrical Impedance Tomography}},
  year = {2023},
  doi = {10.3390/s23042233},
}
Figure