MMV-Net: A Multiple Measurement Vector Network for Multifrequency Electrical Impedance Tomography

Zhou Chen, Jinxi Xiang, Pierre-Olivier Bagnaninchi, Yunjie YangIEEE TNNLS 2023 · OpenReview

TL;DR

This paper introduces MMV-Net, a novel deep learning approach for multi-frequency electrical impedance tomography (mfEIT). By unrolling the MMV-ADMM algorithm and incorporating spatial self-attention and ConvLSTM modules, MMV-Net effectively captures complex frequency-dependent conductivity patterns, achieving superior image quality and faster convergence compared to existing methods.

What Problem Does This Paper Address?

Multi-frequency Electrical Impedance Tomography (mfEIT) is a promising non-invasive imaging technique that leverages how the electrical conductivity of biological tissues changes with frequency to reveal detailed internal structures. However, reconstructing these frequency-dependent conductivity maps is challenging due to the inherent ill-posedness of the inverse problem, low spatial resolution of traditional methods, and high computational costs. Existing deep learning approaches often focus on single-frequency EIT, which is insufficient for the rich information provided by mfEIT. This paper aims to bridge this gap by developing a deep learning framework specifically designed for the complex multi-frequency scenario, addressing the need for accurate, efficient, and robust reconstruction.

Methods

The core of MMV-Net is based on unrolling the iterative steps of the Alternating Direction Method of Multipliers (ADMM) algorithm, specifically tailored for the Multiple Measurement Vector (MMV) model. The MMV model is advantageous because it considers all frequency measurements simultaneously, exploiting correlations between them.

Here's a breakdown of the key methodological innovations:

  • MMV-ADMM Unrolling: The paper takes the standard MMV-ADMM algorithm and transforms its iterative update steps into a deep neural network architecture. This allows for end-to-end learning of parameters, rather than manual tuning.
  • Generalized Non-linear Shrinkage Operator: A crucial part of the MMV-ADMM is a non-linear shrinkage operator that enforces sparsity. In MMV-Net, this operator is replaced by a custom deep network module.
  • Spatial Self-Attention (SSA) Module: This module is designed to capture spatial correlations within individual conductivity images at different frequencies. It helps in identifying structural information and refining feature maps.
  • Convolutional Long Short-Term Memory (ConvLSTM) Module: This module is used to learn the inter-frequency correlations, effectively modeling how conductivity changes across different frequencies. It's particularly useful for capturing the trend of varying conductivity contrasts.
  • Parameter Sharing: To reduce the number of trainable parameters and improve efficiency, parameters are shared across all iterations of the unrolled network.

Datasets

  • Edinburgh mfEIT Dataset: This is a newly introduced dataset specifically for this work. It comprises 4 × 12,414 voltage-conductivity samples generated from simulations of multi-object, multi-conductivity phantoms at four frequencies. The dataset is designed to mimic tissue engineering applications and includes variations in object size, conductivity, and frequency. It is partitioned into training, validation, and testing sets.

Key Insights and Ideas

  • The Power of MMV for mfEIT: The paper strongly advocates for the MMV model over single-measurement vector (SMV) approaches for mfEIT, demonstrating that considering all frequencies simultaneously leads to significantly better reconstruction quality.
  • Deep Learning for Regularization: Replacing the hand-crafted non-linear shrinkage operator with a learned deep network (SSA + ConvLSTM) allows for a more flexible and powerful regularization that can capture complex spatial and frequency dependencies.
  • Synergy of SSA and ConvLSTM: The combination of spatial self-attention and ConvLSTM is crucial. SSA helps understand the structure within each frequency image, while ConvLSTM models how that structure evolves across frequencies, leading to a more holistic reconstruction.
  • Model-Based Deep Learning Advantage: By unrolling a known iterative algorithm (MMV-ADMM), the network inherits the interpretability and physical grounding of model-based methods while benefiting from the learning power of deep networks.

Key Results

  • Superior Image Quality: MMV-Net consistently outperforms state-of-the-art methods like MMV-ADMM, MoDL, and FISTA-Net in terms of PSNR, SSIM, and RMSE across all tested frequencies on the Edinburgh mfEIT Dataset. For example, average PSNR improved from 26.75 dB for FISTA-Net to 28.61 dB for MMV-Net.
  • Faster Convergence: MMV-Net achieves its best performance with significantly fewer iterations (around 7) compared to MMV-ADMM (hundreds of iterations), making it much more computationally efficient.
  • Noise Robustness: The proposed MMV-Net demonstrates better robustness to noise compared to other learning-based methods, maintaining higher image quality even with increased noise levels.
  • Generalization: The method shows good generalization capabilities, performing well on both simulated and real-world experimental data from different EIT sensors and phantom setups.
  • Reduced Parameters: MMV-Net has fewer trainable parameters than some other model-based learning approaches, contributing to its efficiency.

Limitations and Future Work

  • Linearization Assumption: The paper notes that the fundamental MMV model relies on a linearization of the forward problem. This approximation can lead to inaccuracies in estimating conductivity values, especially when dealing with highly non-linear scenarios or when distinguishing very close conductivity values.
  • Training Data: While the proposed MMV-Net is robust to noise in testing, the current training was performed on noise-free data. The authors suggest that adding noisy data to the training set could further enhance its noise robustness.
  • 3D Extension: The current work focuses on 2D mfEIT. The authors propose investigating a 3D version of MMV-Net for more comprehensive imaging applications, such as 3D cell culture imaging.

Summary updated

Abstract

Multifrequency electrical impedance tomography (mfEIT) is an emerging biomedical imaging modality to reveal frequency-dependent conductivity distributions in biomedical applications. Conventional model-based image reconstruction methods suffer from low spatial resolution, unconstrained frequency correlation, and high computational cost. Deep learning has been extensively applied in solving the EIT inverse problem in biomedical and industrial process imaging. However, most existing learning-based approaches deal with the single-frequency setup, which is inefficient and ineffective when extended to the multifrequency setup. This article presents a multiple measurement vector (MMV) model-based learning algorithm named MMV-Net to solve the mfEIT image reconstruction problem. MMV-Net considers the correlations between mfEIT images and unfolds the update steps of the Alternating Direction Method of Multipliers for the MMV problem (MMV-ADMM). The nonlinear shrinkage operator associated with the weighted <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$l_{2,1}$ </tex-math></inline-formula> regularization term of MMV-ADMM is generalized in MMV-Net with a cascade of a Spatial Self-Attention module and a Convolutional Long Short-Term Memory (ConvLSTM) module to better capture intrafrequency and interfrequency dependencies. The proposed MMV-Net was validated on our Edinburgh mfEIT Dataset and a series of comprehensive experiments. The results show superior image quality, convergence performance, noise robustness, and computational efficiency against the conventional MMV-ADMM and the state-of-the-art deep learning methods.
BibTeX
@article{chen2023mmv,
  author = {Zhou Chen and Jinxi Xiang and Pierre-Olivier Bagnaninchi and Yunjie Yang},
  title = {{MMV-Net: A Multiple Measurement Vector Network for Multifrequency Electrical Impedance Tomography}},
  journal = {IEEE Trans. Neural Networks Learn. Syst.},
  year = {2023},
  url = {https://openreview.net/forum?id=JDBbbfGkkB},
}
Figure