Solving Electrical Impedance Tomography with Deep Learning

Yuwei Fan, Lexing Ying

Abstract: This paper introduces a new approach for solving electrical impedance tomography (EIT) problems using deep neural networks. The mathematical problem of EIT is to invert the electrical conductivity from the Dirichlet-to-Neumann (DtN) map. Both the forward map from the electrical conductivity to the DtN map and the inverse map are high-dimensional and nonlinear. Motivated by the linear perturbative analysis of the forward map and based on a numerically low-rank property, we propose compact neural network architectures for the forward and inverse maps for both 2D and 3D problems. Numerical results demonstrate the efficiency of the proposed neural networks.

Introduction

Electrical Impedance Tomography (EIT) is a non-invasive imaging technique that estimates the internal electrical conductivity of an object based on voltage and current measurements taken at its boundary. Because it is radiation-free and inexpensive, it is an ideal candidate for applications ranging from monitoring lung function in medical settings to inspecting industrial pipelines.

Mathematically, EIT is known as the Calderón problem, which involves two primary mappings:

  1. The Forward Map: Predicting boundary measurements given a known internal conductivity.
  2. The Inverse Map: Reconstructing the internal conductivity from boundary measurements.

Historically, the inverse problem has been notoriously difficult because it is severely ill-conditioned. Tiny errors in measurement noise can lead to massive, nonsensical errors in the reconstructed image. Traditional iterative solvers are computationally expensive and require carefully hand-tuned regularization to remain stable.

Methods: The BCR-Net Architecture

Rather than treating EIT as a generic "black box" image-to-image problem, the authors utilize perturbative analysis—studying how small changes in conductivity affect boundary data—to inform their network design. They demonstrate that the mapping between the internal conductivity and the boundary data can be effectively reduced to a map between lower-dimensional functions.

Neural network architecture for the inverse map, featuring BCR-Net 1D and a 2D CNN refinement stage.

The core of their approach is the BCR-Net, a multiscale architecture inspired by the "nonstandard form" of wavelet decomposition. This design is particularly efficient at handling global, translation-invariant operators. The architecture follows a three-module structure:

  1. Encoding: Compresses high-dimensional 2D or 3D data into a set of 1D or 2D functions.
  2. Intermediate (BCR-Net): Performs the primary mapping using multiscale convolutions.
  3. Decoding: Expands the processed data back into the full dimensionality of the output. For the inverse map, an additional 2D CNN stage is used to refine the final reconstruction.

Key Insights: Detection Configurations

A critical physical limitation addressed in the paper is the difference between "one-sided" and "two-sided" detection. In one-sided detection, electrodes are placed only on one boundary (e.g., the top), which works well for shallow objects but fails to resolve deep interior structures. Two-sided detection, placing electrodes on both top and bottom, is necessary for full interior resolution. The proposed architecture handles these variations by concatenating data from different boundary segments.

One-sided vs. two-sided detection setups in EIT, illustrating how boundary measurements relate to internal conductivity.

Key Results and Robustness

The study utilized synthetic datasets created by numerically solving the governing partial differential equations. Testing included both "smooth potentials" (Gaussian blobs) and "shape reconstruction" (triangles and squares).

  • Efficiency: The 2D networks are remarkably compact, using only $10^4 \sim 10^5$ parameters, while 3D versions use $10^5 \sim 10^6$.
  • Accuracy: The forward problem achieved relative errors as low as 0.39%. The inverse problem, while harder, achieved relative errors around 6% for shallow inclusions.
  • Robustness: The model successfully reconstructed shapes even with 1% measurement noise. While boundaries became blurrier as noise increased, the network's learned "prior" provided automatic regularization that outperformed many traditional methods.
  • Speed: Once trained, the neural network produces reconstructions almost instantaneously, a significant improvement over iterative methods that require hundreds of cycles.

Conclusion and Future Work

While the current work demonstrates high efficiency on rectangular and cuboid domains, real-world applications like human torso imaging involve complex, irregular geometries. Future research will focus on reparameterizing arbitrary convex domains into standard shapes and investigating the model's vulnerability to adversarial perturbations. Nonetheless, the physics-informed compression used in BCR-Net represents a significant step forward in making EIT a more practical and robust imaging tool.

Summary updated

Abstract

This paper introduces a new approach for solving electrical impedance tomography (EIT) problems using deep neural networks. The mathematical problem of EIT is to invert the electrical conductivity from the Dirichlet-to-Neumann (DtN) map. Both the forward map from the electrical conductivity to the DtN map and the inverse map are high-dimensional and nonlinear. Motivated by the linear perturbative analysis of the forward map and based on a numerically low-rank property, we propose compact neural network architectures for the forward and inverse maps for both 2D and 3D problems. Numerical results demonstrate the efficiency of the proposed neural networks.
BibTeX
@article{fan2020solving,
  author = {Yuwei Fan and Lexing Ying},
  title = {{Solving Electrical Impedance Tomography with Deep Learning}},
  journal = {Journal of Computational Physics},
  year = {2020},
  url = {https://openreview.net/forum?id=IEo_tuArUe},
}

Loading related content...

Figure