PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Nesterov Momentum Method for Electrical Impedance Tomography with the Complete Electrode Model

Kai Zhu, Jijun Liu, Min Zhong

arXiv:2608.25837v1math.NA

TL;DR

The paper addresses nonlinear, ill-posed conductivity reconstruction from noisy CEM-EIT measurements. It applies adaptive Nesterov momentum with dual-to-primal structural penalties and evaluates the method on measured KIT4 data, where TV-type penalties generally provide more coherent localization while small, closely spaced interior targets remain difficult to resolve.

  • Problem

    Stable and computationally efficient nonlinear reconstruction from noisy electrode data remains challenging in the finite-measurement CEM-EIT inverse problem.

  • Method

    Adaptive Nesterov momentum combines adjoint-gradient correction, momentum extrapolation, dual-to-primal reconstruction, structural penalties, and Sobolev-smoothed gradients for nonlinear CEM-EIT inversion.

  • Results

    Smoothed TV and Huber-TV provide more coherent localization and broadly similar behavior across tested KIT4 configurations, while L2 is diffuse and L1 can geometrically deform supports.

  • Takeaways & Limitations

    The adaptive Nesterov framework can be applied effectively to measured CEM-EIT data with different structural penalties, with TV-type penalties generally offering the best reconstruction balance.

  • Takeaways & Limitations

    Small, closely spaced interior inclusions are not reliably resolved, producing weak, diffuse, or overlapping responses across the penalties.

Abstract

from arXiv · show

We apply the adaptive Nesterov momentum (ANM) method [30] to electrical impedance tomography under the complete electrode model. The forward problem is formulated in the variational CEM setting, accounting for finite electrode size, contact impedance, insulating gaps, and the mean-free voltage gauge. The resulting nonlinear inverse problem is treated within a unified dual-to-primal framework using three classes of strongly convex structural penalties: an L2-type penalty, an L1-type penalty promoting sparse deviations from a calibrated homogeneous background, and a TV-type penalty favoring approximately piecewise-constant conductivities with sharp interfaces. The TV class is implemented using both smoothed TV and Huber-TV formulations. The data- misfit gradient is computed through CEM adjoint equations and stabilized by Sobolev smoothing. The method is evaluated on the publicly available KIT4 tank measurement data after calibration of the background conductivity and contact impedance from no-object measurements. The experiments include single, multiple, mixed-conductivity, and geometrically challenging phantom configurations. The L2-type penalty generally produces smooth but diffuse reconstructions, whereas the L1-type penalty yields cleaner backgrounds with occasional geometric distortion. The smoothed TV and Huber-TV penalties provide more spatially coherent localization and exhibit similar reconstruction behavior across most tested configurations. These results demonstrate the practical applicability of the adaptive Nesterov framework to measured CEM-EIT data.

1 Introduction

The paper applies adaptive Nesterov momentum to nonlinear CEM-EIT reconstruction and evaluates structural penalties on measured KIT4 data. It compares L2-, L1-, and TV-type formulations, including smoothed TV and Huber-TV, with Sobolev-smoothed adjoint gradients.

  • The complete electrode model represents finite electrodes, insulating gaps, electrode–medium contact impedances, and grounded electrode voltages.
  • Stable and computationally efficient nonlinear reconstruction from noisy electrode data remains challenging.
  • Adaptive Nesterov momentum accelerates Landweber-type reconstruction by combining dual-variable gradient updates, momentum extrapolation, and dual-to-primal conductivity recovery.
  • The method evaluates L2-, L1-, and TV-type structural penalties on measured KIT4 data, using smoothed TV and Huber-TV variants.
  • L2-type penalties produce smooth but diffuse variations, while L1-type penalties promote sparse deviations from a calibrated homogeneous background and TV favors sharper, coherent interfaces.

2 EIT with the Complete Electrode Model

The CEM formulates conductivity reconstruction from finite-electrode voltage measurements as a nonlinear, ill-posed inverse problem. The paper defines the forward map and uses derivative and adjoint information within iterative regularization.

  • The CEM models a bounded conductive domain with positive conductivity, finite disjoint electrodes, and positive contact impedances.
  • The CEM determines the interior electric potential and electrode voltages for each prescribed, charge-conserving current pattern.
  • A grounding condition removes the additive-constant ambiguity in the electric potential and electrode voltages.
  • The contact impedance is estimated beforehand and held fixed during conductivity reconstruction, defining the forward map used in inversion.
  • The finite-measurement inverse problem recovers conductivity from noisy electrode-voltage data and is nonlinear, ill-posed, and regularized iteratively.

3 Adaptive Nesterov Momentum Method

The adaptive Nesterov method accelerates nonlinear CEM-EIT reconstruction through momentum, dual-to-primal updates, structural penalties, and Sobolev-smoothed adjoint gradients.

  • Adaptive Nesterov momentum: The method uses an extrapolated dual variable so information from two consecutive iterations accelerates iterative conductivity reconstruction.Each iteration combines adjoint-gradient correction, Nesterov momentum extrapolation, and dual-to-primal reconstruction.
  • Dual-to-primal formulation: Strong convexity enables conductivity recovery from the dual variable while incorporating structural penalties and pointwise admissibility constraints.The dual-to-primal map is the component that differs across L2-, L1-, and TV-type reconstructions.
  • Gradient stabilization: Sobolev smoothing replaces the raw adjoint gradient to suppress small-scale oscillations in update directions.The smoothed gradient is defined using a positive smoothing parameter q.
  • Structural penalty functionals: The L2 penalty provides a smooth quadratic baseline with an explicit pointwise projection update.Its strong-convexity constant is κΘ = 1/2 under the stated convention.
  • Structural penalty functionals: TV penalties favor piecewise-constant conductivities with sharp interfaces but require spatially coupled denoising subproblems.Smoothed TV uses a differentiable approximation and PD–NT, while Huber-TV smooths small gradients while retaining linear growth for larger gradients.
  • Structural penalty functionals: The L1 penalty promotes localized deviations from a calibrated homogeneous background and admits an explicit soft-thresholding and projection update.Its quadratic term ensures strong convexity, while pointwise separability makes the dual-to-primal problem explicit.

4 Validation on Measured KIT4 Data

The adaptive Nesterov framework is evaluated on measured KIT4 tank data using L2-, L1-, and TV-type structural penalties, with TV implemented in smoothed and Huber forms.

  • Validation on measured KIT4 data: Measured KIT4 data are reconstructed within one CEM-based adaptive Nesterov framework using L2-, L1-, and TV-type penalties.The TV penalty is implemented using both smoothed TV and Huber-TV formulations.

4.1 KIT4 data set and computational details

The experiments use publicly available KIT4 measurements from a 16-electrode saline tank, with finite-element CEM discretization and calibration from no-object data.

  • KIT4 data set: The KIT4 data were acquired in a 28 cm cylindrical saline tank equipped with 16 rectangular boundary electrodes.The experiments use N = 16 adjacent current-injection patterns with 2 mA amplitude.
  • Computational details: The CEM and adjoint problem were discretized on 7,132 triangular elements with 3,695 vertices.The mesh was locally refined near boundary electrodes to resolve electrode geometry and interface terms.
  • Calibration: Background conductivity and contact impedance were calibrated from no-object measurements before reconstructing inclusions.The calibrated background was used in the L1 penalty, while the estimated contact impedance remained fixed in later forward solves.
  • Calibration: The no-object reconstruction is nearly spatially uniform and is reasonably consistent with the experimentally reported saline conductivity of 300 µS/cm at 19 °C.Subsequent reconstructions are converted to three-dimensional bulk conductivity and displayed in µS/cm.

Numerical setting

The numerical setting fixes algorithmic, noise, stopping, and smoothing parameters, while selecting penalty parameters empirically from a preliminary study.

  • Numerical setting: The implementation initializes ζ0(x) ≡ 1 and sets µ0 = 1.9κΘ, µ1 = 600, bλn = n/(n + 3), and q = 10^-2.These values summarize the initial dual variable, step-size parameters, momentum bound, and Sobolev smoothing parameter.
  • Numerical setting: The estimated tangential cone constant is η = 0.25, the noise level is δ = 0.0314, and the discrepancy parameter is τ = 1.75.The chosen τ satisfies the required condition cη,τ > 0.
  • Numerical setting: The maximum number of outer iterations is kmax = 800.The stopping and regularization parameters are specified for the measured-data experiments.
  • Numerical setting: Penalty parameters are selected empirically using a preliminary study balancing background suppression, target contrast, and geometric fidelity.The paper states that no theoretical parameter-choice rule for β is available in the measured-data setting.
  • Numerical setting: The smoothed-TV and Huber-TV parameters are fixed at ε = 10^-6 and γ = 10^-3.These values are used in all subsequent experiments.

4.2 Reconstruction results

Across measured CEM-EIT reconstructions, all four penalties identify principal conductivity perturbations in most configurations, but differ in smoothness, background homogeneity, geometric fidelity, and separation of nearby inclusions. TV-type penalties generally provide the most spatially coherent localization, while difficult small interior targets remain unresolved.

  • Each figure presents the reference configuration alongside L2-, L1-, smoothed-TV-, and Huber-TV-type ANM reconstructions, with red contours marking approximate target boundaries.
  • The L2-type penalty locates targets but produces broad, diffuse conductivity responses and contrast spreading.This limits interface sharpness and target-extent resolution, especially for smaller or more distant inclusions.
  • The L1-type penalty yields cleaner, more homogeneous backgrounds but can distort supports, fragment targets, or create artificial bridges between nearby inclusions.These effects occur across single, multiple, and mixed-conductivity configurations.
  • Smoothed TV and Huber-TV generally provide sharper, more coherent localization, preserve relative target arrangements, and behave similarly across configurations.Their reconstructions can still smooth polygonal boundaries and retain boundary artifacts in complicated cases.
  • All four penalties lose resolving power for small, closely spaced targets away from the electrodes, which may appear only as weak, diffuse, or overlapping structures.The individual sizes and boundaries of such targets cannot be reliably recovered.

4.3 Summary of numerical findings

The adaptive Nesterov framework supports measured CEM-EIT reconstruction with multiple structural penalties, whose reconstruction characteristics differ in smoothness, localization, geometry, and inclusion resolution.

  • The framework can be applied to measured CEM-EIT data using L2-, L1-, smoothed-TV-, and Huber-TV-type structural penalties.The experiments use measured KIT4 data and compare these penalty classes across configurations.
  • L2-type reconstructions are smooth but generally spatially diffuse, while background-centered L1 reconstructions suppress weak variations but may show geometric deformation or fragmentation.
  • Smoothed TV and Huber-TV provide more coherent localization and better preserve the relative arrangement of multiple inclusions, with broadly similar behavior across tested configurations.
  • Large or boundary-adjacent inclusions are generally identified more clearly, whereas small interior or closely spaced same-sign targets produce weaker or overlapping responses.Opposite-sign conductive and insulating targets remain more readily distinguishable in mixed configurations.

A Parameter Selection for the Strong Convexity Weight

The strong-convexity weight β controls the balance between the quadratic term and structural penalty, so it is selected empirically to balance background suppression, contrast, and geometric fidelity.

  • β controls the relative contribution of the quadratic strong-convexity term and structural penalty in the dual-to-primal minimization.Its choice affects background suppression, conductivity contrast, and reconstructed-support geometry.
  • Because no theoretical parameter-choice rule is available for this measured CEM-EIT setting, β is selected empirically through preliminary studies over candidate values.The tested ranges are penalty-specific for L1, smoothed TV, and Huber-TV formulations.
  • For L1, increasing β localizes responses and suppresses background variation but can increase support fragmentation or geometric deformation; β = 10 gives strong suppression with more pronounced fragmentation.βL1 = 5 is chosen as a compromise among background homogeneity, target contrast, and geometric representation.
  • Smoothed-TV reconstructions are comparatively less sensitive to β: candidate values recover principal perturbations, while intermediate values yield more coherent target regions.
  • Huber-TV shows similar robustness, with gradual pattern changes across β; increasing β sharpens dominant responses but excessive values can reduce geometric smoothness.The selected value is βTVγ = 2.
  • Overall, L1 is more sensitive to β, whereas smoothed TV and Huber-TV remain stable over a broader parameter range.The final choices are based on qualitative balance among background suppression, target contrast, and geometric fidelity.
Loading 2608.25837v1…