Unsupervised learning methods, such as Deep Image Prior (DIP), have shown great potential in engineering imaging due to their training-data-free nature and high generalization capability. However, their reliance on numerous network parameter iterations results in high computational costs, limiting their practical application, particularly in complex three-dimensional (3D) or time-sequence tomographic imaging tasks. To overcome these challenges, we propose Deep Dynamic Image Prior (D 2 IP), a novel framework for three-dimensional time-sequence imaging. D 2 IP introduces three key strategies — Unsupervised Parameter Warm-Start (UPWS), Temporal Parameter Propagation (TPP), and a customized lightweight reconstruction backbone, Three-dimensional Fast Residual U-Net (3D-FastResUNet) — to accelerate convergence, enforce temporal coherence, and improve computational efficiency. Experimental results on both simulated and clinical pulmonary datasets demonstrate that D 2 IP enables fast and accurate 3D time-sequence Electrical Impedance Tomography (tsEIT) reconstruction. Compared to the state-of-the-art Regularized Shallow Image Prior (R-SIP) baseline, D 2 IP delivers superior image quality — with a 24.8% increase in average Mean Structural Similarity Index (MSSIM) and an 8.1% reduction in Relative Error (ERR) — alongside significantly reduced computational time (7.1× faster), demonstrating its promise for artificial intelligence (AI)-driven medical engineering applications, as exemplified by clinical dynamic pulmonary imaging.

Multi-frequency Electrical Impedance Tomography (mfEIT) has demonstrated considerable potential for non-invasive, real-time monitoring of physiological states by reconstructing tissue conductivity at multiple frequencies. However, practical deployment in Internet-of-Things (IoT)-enabled medical applications, such as wearable thoracic sensors, remains challenging due to the heavy reliance on large-scale datasets and poor generalization of current methods. To address this, we present a model-based, dataset-free unsupervised mfEIT reconstruction framework, termed Multi-Branch Attention Image Prior (MAIP). Our approach leverages a specially designed Multi-Branch Attention Network (MBA-Net) to iteratively reconstruct conductivity images at multiple frequencies without the need for training data. By implicitly regularizing the reconstruction process, MBA-Net effectively captures inter- and intra-frequency correlations, ensuring accurate and robust imaging suitable for resource-limited IoT scenarios. Extensive simulations and real-world experiments validate that the proposed method surpasses state-of-the-art approaches in reconstruction accuracy and generalization ability, thus providing a promising solution for IoT-based medical imaging applications.

The inherent compliant nature of soft robots can offer remarkable advantages over their rigid counterparts in terms of safety to human users and adaptability in unstructured environments. However, this feature also magnifies the complexity of their bodies, rendering their proprioception, and hence their control, extremely challenging. Given this intricacy, machine learning is a potent candidate for extracting proprioceptive insights from sensor data due to its proven capabilities in tackling analogous issues in computer vision (CV) and natural language processing (NLP). Recently, key aspects of soft robot proprioception have been addressed via learning-based techniques, but most of these are rooted in the supervised learning (SL) paradigm. This typically requires collecting a large number of costly annotated training samples, thereby constraining its widespread and speedy adoption in real-world applications. To mitigate this limitation, we propose a self-SL framework for soft robot proprioception. Our method utilizes vast unannotated data for network pretraining by self-SL. Then, the pretrained model is fine-tuned with a limited set of annotated samples by SL. We validate the proposed method's efficacy on a high-resolution 3-D morphological reconstruction task using a publicly available dataset. Remarkably, our approach is shown to necessitate only about 1/20 of annotated samples to achieve better performance than the fully supervised method.

With the increasing demand for human-computer interaction (HCI), flexible wearable gloves have emerged as a promising solution in virtual reality, medical rehabilitation, and industrial automation. However, the current technology still has problems like insufficient sensitivity and limited durability, which hinder its wide application. This paper presents a highly sensitive, modular, and flexible capacitive sensor based on line-shaped electrodes and liquid metal (EGaIn), integrated into a sensor module tailored to the human hand’s anatomy. The proposed system independently captures bending information from each finger joint, while additional measurements between adjacent fingers enable the recording of subtle variations in inter-finger spacing. This design enables accurate gesture recognition and dynamic hand morphological reconstruction of complex movements using point clouds. Experimental results demonstrate that our classifier based on Convolution Neural Network (CNN) and Multilayer Perceptron (MLP) achieves an accuracy of 99.15% across 30 gestures. Meanwhile, a transformer-based Deep Neural Network (DNN) accurately reconstructs dynamic hand shapes with an Average Distance (AD) of 2.076±3.231 mm, with the reconstruction accuracy at individual key points surpassing SOTA benchmarks by 9.7% to 64.9%. The proposed glove shows excellent accuracy, robustness and scalability in gesture recognition and hand reconstruction, making it a promising solution for next-generation HCI systems.

Hyper-redundant tendon-driven manipulators offer greater flexibility and compliance over traditional manipulators. A common way of controlling such manipulators relies on adjusting tendon lengths, which is an accessible control parameter. This approach works well when the kinematic configuration is representative of the real operational conditions. However, when dealing with manipulators of larger size subject to gravity, it becomes necessary to solve a static force problem, using tendon force as the input and employing a mapping from the configuration space to retrieve tendon length. Alternatively, measurements of the manipulator posture can be used to iteratively adjust tendon lengths to achieve a desired posture. Hence, either tension measurement or state estimation of the manipulator are required, both of which are not always accurately available. Here, we propose a solution by reconciling cables tension and length as the input for the solution of the system forward statics. We develop a screw-based formulation for a tendon-driven, multi-segment, hyper-redundant manipulator with elastic joints and introduce a forward statics iterative solution method that equivalently makes use of either tendon length or tension as the input. This strategy is experimentally validated using a traditional tension input first, subsequently showing the efficacy of the method when exclusively tendon lengths are used. The results confirm the possibility to perform open-loop control in static conditions using a kinematic input only, thus bypassing some of the practical problems with tension measurement and state estimation of hyper-redundant systems.

A novel unsupervised algorithm, named Deep Learning-based Error Image Prior (DLEIP), is proposed for lung Electrical Impedance Tomography (EIT). An Attention-guided Denoising Network (AD-Net) is employed in DLEIP algorithm to optimizes the initial conductivity distribution through built-in back-propagation. The sparse block, feature enhancement block and attention block are also introduced to AD-Net for improving learning efficiency, while incorporating residual learning technique to extract the deep error image prior. The simulation results indicate that the ventilation areas and the lesions can be effectively reconstructed by DLEIP algorithm. The Correlation Coefficients (CCs) of the reconstructed images are higher than 0.82, and the Relative Errors (REs) are lower than 0.40. Moreover, the DLEIP algorithm has good regularization integration ability. By integrating with Total Variation (TV) regularization, the DLEIP-TV algorithm can be obtained to further improves the imaging accuracy. During the experiment, the mapping model is constructed in the circular domain to verify the practicality of the proposed algorithm. The results indicate that the shape and size of the object in the domain can be reconstructed more accurately. The CCs of the reconstructed images are higher than 0.85, and the REs are lower than 0.26.

WiFi fingerprinting has been extensively studied for years to provide location estimation in indoor scenarios. Researchers have used various machine learning algorithms to match the online fingerprint to the pre-collected offline fingerprints, which have location labels, for location estimation. However, neither conventional machine learning algorithms nor modern deep neural networks explore the geometric relations between WiFi access points and the location where the fingerprint was taken. Therefore, they cannot capture the non-Euclidean nature of the WiFi fingerprint. Furthermore, prior research has indicated that fusing multiple modalities can improve positioning performance compared to using only WiFi. Therefore, this study proposes a novel Multimodal Graph Fingerprinting method for indoor positioning systems. The proposed method constructs a multimodal graph at the location of the user’s smart terminal by fusing radio frequency signals, electromagnetic field (EMF) strength, and inertial sensor measurements. A hierarchical deep graph neural network is developed to learn the relations between the multimodal graphs and their locations by capturing the features of the identities (such as MAC addresses, WiFi Received Signal Strength (RSS), and EMF data) and the topology information. Experiments on a real dataset built on a university campus demonstrate that the proposed model can achieve a median positioning error of 2.1m by fusing different modalities.

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$l_{2,1}$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.

WiFi received signal strength (RSS)-based finger-printing has been widely adopted in many indoor positioning systems due to its implementation simplicity and low computational complexity. Recently, some studies have explored the potential temporal features among WiFi data to provide better positioning accuracy using Long Short-Term Memory (LSTM). However, the large volume of invalid RSS signals/values in the radio map degrades the training performance and the positioning accuracy. Some recent research has shown effort to reduce the impact of the problem mentioned above using deep learning. However, these either lack efficiency (repeated training needed or training resource wasted) or cannot provide precision position estimations. This paper proposes a novel Encoded LSTM network model to efficiently extract main features from the WiFi RSS data and provide high positioning accuracy. Experimental results based on an open-source crowdsourced dataset show that the encoded LSTM can reduce the dimension of the WiFi fingerprints from 992 to 64 and achieve a mean positioning error of 7.37m. Compared to the benchmark results based on the same dataset, the encoded LSTM shows the lowest mean error, which outperforms 13 popular positioning algorithms. The proposed encoded LSTM can also provide a 10% improvement in positioning accuracy in comparison to other conventional LSTM models.

WiFi received signal strength (RSS)-based finger-printing has attracted much attention in indoor positioning in the past decade. One WiFi fingerprint comprises multiple RSS values annotated with the location (reference point) where the signals are obtained. The positioning accuracy of WiFi RSS fingerprinting-based indoor positioning systems is highly reliant on the data volume of the observed signals. However, taking such data in a large complex indoor area is usually time-consuming and labor-intensive. In recent years, crowdsourcing approaches have been proposed to collect WiFi data and record the location by utilizing the trajectories of common users to reduce the burden of constructing the database. Nevertheless, crowdsourced data is usually sensitive to crowd density. The data coverage is usually not enough to cover the entire targeted environment to provide good positioning accuracy, particularly at the beginning stage of constructing a database. Besides, it is also expected that some regions do not have enough fingerprints to provide good positioning performance since they do not have as many visitors as others. Therefore, this paper proposes a WiFi fingerprint augmentation method to generate more fingerprints by predicting RSS values on unsurveyed locations through a multivariate Gaussian process regression (MGPR) model. Evaluations are conducted on an open-source crowdsourced WiFi fingerprint dataset collected in an actual multi-floor university building. The experiment results show that the proposed WiFi fingerprint augmentation method can enhance the global data coverage (considering the entire building) to reduce the positioning error by 5% to 20%. Also, the proposed method can sharply reduce the positioning error in some indoor regions by improving local data density (considering a 2D region on a certain floor).

While Electrical Impedance Tomography (EIT) has found many biomedicine applications, better image quality is needed to provide quantitative analysis for tissue engineering and regenerative medicine. This paper reports an impedance-optical dual-modal imaging framework that primarily targets at high-quality 3D cell culture imaging and can be extended to other tissue engineering applications. The framework comprises three components, i.e., an impedance-optical dual-modal sensor, the guidance image processing algorithm, and a deep learning model named multi-scale feature cross fusion network (MSFCF-Net) for information fusion. The MSFCF-Net has two inputs, i.e., the EIT measurement and a binary mask image generated by the guidance image processing algorithm, whose input is an RGB microscopic image. The network then effectively fuses the information from the two different imaging modalities and generates the final conductivity image. We assess the performance of the proposed dual-modal framework by numerical simulation and MCF-7 cell imaging experiments. The results show that the proposed method could improve the image quality notably, indicating that impedance-optical joint imaging has the potential to reveal the structural and functional information of tissue-level targets simultaneously.

Inverse problems are essential to imaging applications. In this letter, we propose a model-based deep learning network, named FISTA-Net, by combining the merits of interpretability and generality of the model-based Fast Iterative Shrinkage/Thresholding Algorithm (FISTA) and strong regularization and tuning-free advantages of the data-driven neural network. By unfolding the FISTA into a deep network, the architecture of FISTA-Net consists of multiple gradient descent, proximal mapping, and momentum modules in cascade. Different from FISTA, the gradient matrix in FISTA-Net can be updated during iteration and a proximal operator network is developed for nonlinear thresholding which can be learned through end-to-end training. Key parameters of FISTA-Net including the gradient step size, thresholding value and momentum scalar are tuning-free and learned from training data rather than hand-crafted. We further impose positive and monotonous constraints on these parameters to ensure they converge properly. The experimental results, evaluated both visually and quantitatively, show that the FISTA-Net can optimize parameters for different imaging tasks, i.e. Electromagnetic Tomography (EMT) and X-ray Computational Tomography (X-ray CT). It outperforms the state-of-the-art model-based and deep learning methods and exhibits good generalization ability over other competitive learning-based approaches under different noise levels.

Imaging the bio-impedance distribution of the brain can provide initial diagnosis of acute stroke. This paper presents a compact and non-radiative tomographic modality, i.e. multi-frequency Electromagnetic Tomography (mfEMT), for the initial diagnosis of acute stroke. The mfEMT system consists of 12 channels of gradiometer coils with adjustable sensitivity and excitation frequency. To solve the image reconstruction problem of mfEMT, we propose an enhanced Frequency-Constrained Sparse Bayesian Learning (FC-SBL) to simultaneously reconstruct the conductivity distribution at all frequencies. Based on the Multiple Measurement Vector (MMV) model in the Sparse Bayesian Learning (SBL) framework, FC-SBL can recover the underlying distribution pattern of conductivity among multiple images by exploiting the frequency constraint information. A realistic 3D head model was established to simulate stroke detection scenarios, showing the capability of mfEMT to penetrate the highly resistive skull and improved image quality with FC-SBL. Both simulations and experiments showed that the proposed FC-SBL method is robust to noisy data for image reconstruction problems of mfEMT compared to the single measurement vector model, which is promising to detect acute strokes in the brain region with enhanced spatial resolution and in a baseline-free manner.

In this paper, we consider the reconstruction of three-dimensional (3-D) conductivity distribution using electrical impedance tomography (EIT) technique. A high-resolution and efficient algorithm is developed to solve the EIT inverse problem. The presented algorithm is extended upon a recently proposed novel EIT reconstruction approach based on structure-aware sparse Bayesian learning (SA-SBL). The correlation between proximal layers in the 3-D geometry are incorporated into the structure prior to improve the reconstruction accuracy. In addition, an efficient approach based on approximate message passing is developed to accelerate the large-scale 3-D learning process. To validate the algorithm, numerical experiments using real recorded data are conducted. The visual and quantitative-metric comparisons show that the proposed method outperforms the existing methods in terms of reconstruction accuracy and computational complexity in all test cases. The SA-SBL-based reconstruction approach can preserve the 3-D structure of medical volume, reduce the systematic artifacts, and improve the computational efficiency.

Electrical impedance tomography (EIT) is developed to investigate the internal conductivity changes of an object through a series of boundary electrodes, and has become increasingly attractive in a broad spectrum of applications. However, the design of optimal tomography image reconstruction algorithms has not achieved the adequate level of progress and matureness. In this paper, we propose an efficient and high-resolution EIT image reconstruction method in the framework of sparse Bayesian learning. Significant performance improvement is achieved by imposing structure-aware priors on the learning process to incorporate the prior knowledge that practical conductivity distribution maps exhibit clustered sparsity and intra-cluster continuity. The proposed method not only achieves high-resolution estimation and preserves the shape information even in low signal-to-noise ratio scenarios but also avoids the time-consuming parameter tuning process. The effectiveness of the proposed algorithm is validated through comparisons with state-of-the-art techniques using extensive numerical simulation and phantom experiment results.