robust eit design

This work addresses the nonlinear, severely ill-posed inverse problem of reconstructing internal conductivity distributions from boundary voltage/current measurements in Electrical Impedance Tomography (EIT), with a specific focus on high-accuracy estimation of the convex hulls of conductivity anomalies. We propose the first end-to-end learnable framework that deeply integrates Ikehata’s enclosure method—a rigorous theoretical approach rooted in complex geometrical optics solutions—with deep neural networks, thereby eliminating reliance on linearized approximations and explicit prior modeling. By jointly leveraging boundary integral equation modeling and supervised network training, our approach synergistically combines theory-driven constraints with data-driven learning. Evaluated on both synthetic and experimental EIT datasets, the method achieves over 30% improvement in convex hull localization accuracy compared to the classical least-squares enclosure method, while demonstrating markedly enhanced robustness against measurement noise and model mismatch.

EIT reconstruction faces two fundamental challenges: ill-posedness of the inverse problem and spatially non-uniform sensitivity. Conventional model-based methods rely on handcrafted regularization but neglect sensitivity heterogeneity, while supervised learning suffers from poor generalizability. Existing neural field approaches lack explicit physical constraints, limiting reconstruction accuracy. To address these issues, we propose an unsupervised, physics-driven neural compensation framework. Our method introduces a novel dynamic neural representation allocation mechanism that adaptively enhances representational capacity in low-sensitivity regions, guided by the sensitivity distribution derived from the EIT forward model. It integrates the physical forward model for implicit regularization and incorporates sensitivity-aware coordinate encoding with adaptive weight modulation. Evaluated on both synthetic and experimental data, our approach significantly improves structural fidelity and artifact suppression—particularly mitigating distortion in low-sensitivity regions—and demonstrates superior robustness compared to state-of-the-art model-based and supervised learning methods.

This study addresses the inverse problem of electrical impedance tomography (EIT) in anisotropic media, aiming to detect and characterize single or multiple heterogeneous inclusions—assessing their existence, quantity, size, and internal conductivity anisotropy—solely from boundary electrostatic measurements (i.e., the Dirichlet-to-Neumann map) acquired via a limited number of electrodes (e.g., 16). We propose a novel hybrid modeling paradigm integrating artificial neural networks (ANNs) and support vector machines (SVMs): ANNs perform regression for inclusion size estimation, while SVMs handle multi-inclusion detection and anisotropy classification. Only two independent boundary excitations are required for high-accuracy prediction. The method eliminates reliance on dense boundary sampling or strong prior models. Validation on both synthetic and experimental data demonstrates superior performance: high inclusion detection accuracy, significantly improved anisotropy classification accuracy, and low size estimation error.

Electrical impedance tomography (EIT) suffers from low spatial resolution when reconstructing high-fidelity conductivity distributions on coarse finite-element meshes. Method: This paper proposes MR-EIT, a dual-modal multi-resolution reconstruction framework supporting both supervised learning and purely voltage-driven unsupervised iterative reconstruction. It introduces a novel co-architecture integrating ordered feature extraction with unordered coordinate encoding, incorporating pretrained feature mapping, symmetric function modeling, local feature aggregation, two-stage joint optimization, and voltage-residual-driven iterative refinement. Contribution/Results: MR-EIT achieves, for the first time in EIT, annotation-free, low-iteration (≥40% reduction), cross-resolution adaptive super-resolution reconstruction. Extensive simulation and tank experiments demonstrate statistically significant improvements over state-of-the-art methods in structural similarity (SSIM) and relative imaging error (RIE), while maintaining superior robustness and reconstruction fidelity under noise.

In 3D high-contrast electrical impedance tomography (EIT), severe interface distortions arise in reconstructions due to strong ill-posedness induced by abrupt conductivity discontinuities—e.g., at metal implant–tissue or defect–material boundaries. To address this, we propose a 3D shape optimization framework based on implicit neural representations (signed distance function networks). Our method introduces, for the first time, a variational optimization scheme integrated with shape derivatives to explicitly encode high-contrast interface conditions. We further design a low-dimensional latent-space implicit shape representation that ensures geometric fidelity while improving parameter efficiency. Theoretical analysis guarantees convergence. Extensive 3D simulations and hardware-in-the-loop experiments demonstrate a 42% reduction in average interface localization error and significantly enhanced reconstruction stability. These results validate the framework’s effectiveness and practical potential for medical monitoring of metallic implants and industrial non-destructive testing.

This work addresses the robust solution of the Calderón inverse conductivity problem in electrical impedance tomography (EIT) under noisy measurements. Methodologically, we propose a noise-aware operator learning framework based on neural operators. Our core innovation lies in extending the original nonlinear inverse operator to a reproducing kernel Hilbert space of integral kernels—preserving mapping stability while enabling efficient neural operator approximation. Specifically, we employ the Fourier neural operator to parameterize the continuous integral kernel and integrate theoretical guarantees from operator approximation under noise perturbations, achieving end-to-end reconstruction of infinite-dimensional piecewise-constant and log-normal conductivity fields. Experiments demonstrate high reconstruction accuracy under strong noise and superior generalization over classical regularization methods. The approach provides a data-driven paradigm for nonlinear inverse problems that is both theoretically grounded—via stability and approximation guarantees—and practically effective.

This work addresses the susceptibility of conductivity reconstruction in electrical impedance tomography (EIT) to noise and its reliance on explicit regularization by proposing an implicit parameterization framework that eliminates the need for such regularization. The method generates physically plausible conductivity distributions through low-dimensional latent variables and a composite nonlinear mapping, incorporating truncated graph Laplacian embeddings to inject structural priors and employing bound-preserving mappings to improve the conditioning of the optimization problem. The approach achieves, for the first time, high-quality three-dimensional time-difference EIT reconstructions, demonstrating superior physical consistency, structural fidelity, and robustness across 2D/3D simulations, phantom experiments, and in vivo lung data. Notably, it substantially enhances 3D spatial resolution, thereby strengthening the clinical potential of EIT for respiratory monitoring.

This work addresses the limitations of conventional tactile sensor arrays—poor scalability, complex wiring, and inadequate conformity to curved surfaces—when deployed over the full body of humanoid robots. The authors propose a flexible, conformal tactile skin based on electrical impedance tomography (EIT), integrating 3D-printed conductive thermoplastic polyurethane (TPU) structures with customizable geometry, a porous sensing layer, and contact-enhancing patches. Contact perception is achieved by measuring boundary voltage changes, and contact locations are efficiently reconstructed using a one-step Gauss–Newton algorithm. The design eliminates the need for robot-specific redesign, enabling scalable, large-area tactile coverage compatible with arbitrary curved surfaces. Validation on planar, U-shaped, and iCub face-mimicking prototypes demonstrates an average localization error of 6 mm across 18 contact points on curved surfaces, without requiring supervised post-processing.

This study addresses the challenges of non-uniform sensitivity and low force reconstruction accuracy in conventional electrical impedance tomography (EIT) for large-area tactile sensing. To overcome these limitations, the authors propose a novel hybrid robotic skin that integrates EIT with pneumatic tactile sensing, fabricated via 3D printing and spray coating to ensure low cost and scalability. By combining Tikhonov regularization-based inverse reconstruction with a single-pad pneumatic calibration strategy, the method significantly reduces sensitivity non-uniformity—evidenced by a decrease in the coefficient of variation from 0.31 to 0.14—and enables consistent, high-fidelity force distribution reconstruction across the entire sensing area. The system’s robust multi-contact perception capability is validated on a humanoid robot’s chest plate, demonstrating its potential as a practical and scalable tactile solution for full-body somatosensory robotics.

研究通过EITWatch智能手表集成平面电导率断层扫描技术，利用8个平面电极采集阻抗变化数据以识别手势，解决了传统腕部EIT系统需额外电极覆盖及独立前端的问题。