Physics-Driven Neural Compensation For Electrical Impedance Tomography

Chuyu Wang, Huiting Deng, Dong Liu · 2025

Read Physics-Driven Neural Compensation For Electrical Impedance Tomography by Chuyu Wang, Huiting Deng, Dong Liu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-18067.txt.

Introduction

Chuyu Wang, Huiting Deng, and Dong Liu present a new way to make images of the inside of the body in 2025. Their method sends mild electric currents through tissue and reads the response. From those reads it builds a map of how well each spot conducts. The task is hard because small errors can blur the map, and some areas respond weakly.

Their method, called PhyNC, uses the physics of the test to guide a neural model. The model gives more capacity to weak areas, so detail there is not lost. Tests on made up data and on real lab data beat older methods. Detail held up better and false marks were fewer, most of all in weak areas.

Catalog record

CREATORChuyu Wang, Huiting Deng, Dong Liu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-18067

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-Driven Neural Compensation For Electrical Impedance Tomography Chuyu Wang, Huiting Deng, Dong Liu Abstract Electrical Impedance Tomography (EIT) provides a non-invasive, portable imaging modality with significant potential in medical and industrial applications. Despite its advantages, EIT encounters two primary challenges: the ill-posed nature of its inverse problem and the spatially variable, location-dependent sensitivity distribution. Traditional model-based methods mitigate ill-posedness through regularization but overlook sensitivity variability, while supervised deep learning approaches require extensive training data and lack generalization. Recent developments in neural fields have introduced implicit regularization techniques for image reconstruction, but these methods typically neglect the physical principles underlying EIT, thus limiting their effectiveness. In this study, we propose PhyNC (Physics-driven Neural Compensation), an unsupervised deep learning framework that incorporates the physical principles of EIT. PhyNC addresses both the ill-posed inverse problem and the sensitivity distribution by dynamically allocating neural representational capacity to regions with lower sensitivity, ensuring accurate and balanced conductivity reconstructions. Extensive evaluations on both simulated and experimental data demonstrate that PhyNC outperforms existing methods in terms of detail preservation and artifact resistance, particularly in low-sensitivity regions. Our approach enhances the robustness of EIT reconstructions and provides a flexible framework that can be adapted to other imaging modalities with similar challenges. August 24, 2026 1 Introduction Non-invasive, portable, and accurate tomographic imaging holds great significance for medical and industrial applications [ 1 ] , [ 2 ] . Though imaging techniques such as computed tomography (CT) can achieve precise imaging [ 3 ] , they are constrained by workspace limitations. In contrast, Electrical Impedance Tomography (EIT) enables convenient, non-invasive imaging by estimating the internal conductivity distribution from boundary voltage measurements [ 4 ] . Owing to these advantages, EIT has found extensive use in various medical contexts, including bioelectronic medicine [ 5 ] , [ 6 ] , medical imaging [ 7 ] , [ 8 ] , [ 9 ] , [ 10 ] , [ 11 ] , and the monitoring of muscle engagement [ 12 ] , [ 13 ] , [ 14 ] . Beyond healthcare, EIT plays an important role in physical human-computer interaction, enabling hand pose estimation [ 15 ] , [ 16 ] , [ 17 ] and the development of soft robotic skin [ 18 ] , [ 19 ] , [ 20 ] , [ 21 ] , [ 22 ] , [ 23 ] , [ 24 ] , [ 25 ] , [ 26 ] . Despite its broad utility, EIT is hindered by two primary challenges [ 2 ] , [ 4 ] : (i) The ill-posed nature of its underlying inverse problem, and (ii) The location-dependent, non-uniform sensitivity distribution [ 27 ] stemming from the properties of the applied current. Traditional model-based methods address the ill-posed nature of inverse problems by constraining the solution space through regularization techniques (e.g., sparse regularization [ 28 ] and total variation regularization [ 10 ] ). Despite these strategies, their modeling capabilities remain limited, because they do not address the second challenge. Recently, deep learning has emerged as a promising alternative to model-based approaches [ 29 ] . Leveraging data-driven priors, supervised deep learning (DL) methods have achieved state-of-the-art performance in ill-posed tomographic imaging [ 30 ] . In the context of EIT, these methods can directly estimate internal conductivity distributions from boundary voltage measurements by training neural networks to learn nonlinear inverse operators [ 31 ] , [ 32 ] , [ 33 ] , [ 34 ] , [ 35 ] , [ 36 ] , [ 37 ] . Nonetheless, these methods require large amounts of training data, which is often scarce or expensive to obtain, and they exhibit limited generalization capabilities since they do not directly tackle EIT’s inherent challenges, relying instead on end-to-end mappings. Neural field approaches represent a novel DL strategy that has shown substantial promise for image reconstruction [ 38 ] , [ 39 ] , [ 40 ] , [ 41 ] , [ 42 ] , [ 43 ] , [ 44 ] , [ 45 ] , [ 46 ] . These methods train a multilayer perceptron (MLP) to represent the reconstructed image as a continuous function that maps the spatial coordinates to the corresponding values. A similar approach, the deep image prior (DIP) [ 47 ] , has also been applied to image reconstruction tasks in EIT [ 48 ] , [ 49 ] as well as in CT [ 50 ] and positron emission tomography (PET) [ 51 ] , [ 52 ] . DIP is based on the assumption that a neural network’s architecture can act as a form of regularization, thereby enabling the solution of inverse problems without requiring training data. Similarly, the neural field enforces intrinsic image consistency, providing a strong implicit regularization that constrains the solution space in an under-determined setting [ 53 ] . Notably, while DIP often suffers from spectral bias [ 53 ] , [ 54 ] , neural field has been shown to mitigate this issue [ 55 ] , thereby improving the capture of high-frequency details. However, despite effectively addressing the ill-posedness of the inverse problem through implicit regularization, existing neural field-based approaches predominantly adhere to computer vision paradigms while neglecting the intrinsic physical properties inherent to the second fundamental challenge of EIT. In this work, we introduce PhyNC ( Phy sics-Driven N eural C ompensation), an unsupervised deep learning framework grounded in EIT’s physical principles. PhyNC addresses both the ill-posed nature of EIT and its non-uniform sensitivity distribution by introducing a neural compensation strategy: it allocates greater representational capacity to coordinates that exhibit lower physical sensitivity, thus balancing the overall sensitivity distribution. By combining the inherent regularization of implicit neural representation with a fidelity constraint derived from the EIT model, PhyNC holds great potential to achieve a more comprehensive and accurate reconstruction than previous approaches. To mitigate the model’s susceptibility to overfitting, we design a hybrid neural representation that synergistically integrates complementary encoding schemes. This approach is further enhanced by frequency-domain regularization and stochastic sampling strategies, which introduce controlled parameter perturbations during optimization to systematically improve robustness and generalization capability. To further evaluate our approach, we developed a Hybhash-based reconstruction method as a comparative technique—introduced here for the first time in the context of EIT—to underscore the advantages of PhyNC. We also provide a detailed analysis of the impact of location-dependent sensitivity on reconstruction quality and examine the effects of finite element (FE) mesh discretization by varying the density of FE nodes and elements. Experimental results on both simulated and real data demonstrate that PhyNC not only reproduces shape details with high fidelity but also robustly resists artifacts across regions of varying sensitivity. Moreover, our analysis reveals that the choice of embedding space is crucial to the success of the proposed method, suggesting that similar neural compensation techniques could have broader applications. Our main contributions are summarized as follows: • We propose PhyNC, an unsupervised deep learning framework that explicitly integrates the physical principles of EIT to address both the inverse problem’s ill-posedness and the non-uniform sensitivity distribution. To the best of our knowledge, this is the first framework of its kind, and it can be readily

Read the complete text