Recent advancements in ultrasound technologies, such as ultrasound localization microscopy and ultrafast ultrasound Doppler, have enabled high-definition imaging of microvasculature. However, detecting weak microbubble or blood flow signals amid strong background noise remains a challenge, particularly in deep tissues. This study aims to enhance the signal contrast of microbubble and blood flow by leveraging their distinct spatial-temporal coherence in comparison to undesired noise for robust microbubble detection and microvascular imaging. We propose to quantify the signal coherence based on similarity analysis of beamformed ultrasound microbubble/blood flow data within the plane wave imaging framework. A spatial pixel is considered more likely to be a true microbubble/blood flow signal with a higher level of similarity, which can be measured by either of the following methods: 1) spatially block-wise normalized cross-correlation between two compounded frames; 2) temporally normalized autocorrelation across multiple compounded frames; 3) normalized cross-correlation between two subsets of post-compounded frames; 4) normalized autocorrelation of the pre-compounded data across angular direction. The original microbubble/blood flow signal is then weighted by the similarity measurement on a pixel-by-pixel basis to generate images with an improved signal contrast. The robustness of the proposed methods was first demonstrated in both phantom experiments and in vivo microbubble data from kidney transplant. We further validated their feasibility in blood flow imaging without the use of microbubbles based on in vivo data of human liver and kidney. Significant contrast improvement was observed, facilitating better visualization and detection of both microbubble and noncontrast microflow signals, which indicates a great potential of the methods for improved microvascular imaging and widespread clinical translation.

This paper describes the historical trajectory of video conferencing systems, spanning from their earlier mechanical and analog origins in the 1920s to the sophisticated IP-based services delivered from the cloud in the 2020s. Each technological age is examined, highlighting the technical and functional aspects that characterized its evolution. Commercial landmarks of each age are presented providing a comprehensive overview of the most prominent offerings at pivotal moments in the timeline. By examining the past and speculating on the future, this paper aims to provide a holistic understanding of the development, current state, and forthcoming trends in video conferencing technology.

Multilevel converters have enabled various applications that are not possible with conventional two-level converters. Many of these applications, however, need a high output bandwidth, often approaching the switching or tolerated loss limit of the transistors, and still high quality, e.g., to actively stabilize and dampen a DC grid or specifically excite certain molecules or neural circuits in medical applications. Modulation in multilevel converters has two dimensions for improving the output quality, namely temporal switching modulation and amplitude quantization. A high bandwidth approaching the switching rate challenges existing modulation methods: carrier-based switching modulation is fine at low frequencies but experiences interaction between the carrier and the signal at the upper end of the spectrum; fundamental-frequency switching, such as nearest-level modulation (NLM), perform well at high frequencies but cause intolerable distortion for low frequency contents. We propose a hybrid modulation concept that can combine any methods from these two classes. It passes the error of a fundamental frequency method through a filtered switching modulator to combine the high output quality of the latter with the high bandwidth of the former. We optimize the filter to avoid under-modulation of the signal with the carrier of the modulator and to achieve the minimum overall distortion throughout a wide output bandwidth. We demonstrate the performance experimentally with a cascaded-bridge converter and compare it with the best prior arts. This technique ensures a usable output bandwidth up to 100% of the switching rate and maintains a total distortion level below 3%.Â

The growing trend of solitary living among the elderly and young, coupled with the high risk of falls leading to injuries and death, highlights the need for fall monitoring systems. Emphasizing individuals' privacy and comfort, these systems should rely on radar sensors instead of visual-based, acoustic-based, or wearable solutions. Current radar-based systems are yet to reach satisfactory real-world performance. This work proposes a radar-based fall detection system with superior performance in complex real-world scenarios while maintaining edge computing capabilities and minimum hardware resources. The proposed deep learning system achieved a recall of 98.99% and a precision of 99.32%. These unprecedented performance numbers are measured on the proposed dataset, which is the most real-life representative dataset in the literature. The system has 211.8k parameters and ~8.84 M Floating Point Operations (FLOPs), achieving an edge computing capability. Moreover, the efficient model construction eliminates redundant computation in real-time operation. Furthermore, this work proposes a novel metric that encompasses all dataset's quality aspects into a single number, which can be applied to all classification problems. This metric can then be used as a correction factor for performance metrics to put them in the context of the dataset used for testing.

As algorithms get better at their accuracy and computational efficiency, they invoke curiosity among the affected scientific communities to check if they can benefit from newer versions or not. Computer vision is one such domain that has observed rapid growth in terms of algorithmic advancements. The advent of deep learning was itself a catalyst for agile innovation. Coupled with rapid improvements in algorithms for object detection, the speed of innovation has become tremendous. And so there are many engineering and scientific disciplines that leverage object detection, any improvement in the latter has a ripple effect on the erstwhile. Computer Aided Laparoscopy (CAL) has come a long way due to object detection algorithms based on deep learning. Yet, every once in a while a new algorithm is released. It is tempting to see how a new algorithm may have affected the tool detection accuracy and efficiency in CAL. Recently version 8 of the famous You Look Only Once (YOLO) algorithm was released. Like all the past releases, it has been claimed that this version is better at detection accuracy as well as computational efficiency. This paper examines the performance of YOLOv8 at tool detection in a CAL context. We employed a well-known laparoscopy tool detection benchmark dataset in this research. Models with superior performance have been obtained as a result of this research. Models are superior in terms of both detection accuracy as well as inference speed. Moreover, the models are ready to be deployed into a production environment. The results that are reported in this paper are not only useful for the surgical community but also for the benchmarking of the YOLO algorithm.

The article introduces a novel channel attention architecture embedded within a sensor fusion framework for fetal movement monitoring. Our proprietary multi-sensory device recorded the training dataset, comprising accelerometric sensor data collected from forty-four pregnant mothers. The channel attention architecture, LBCAM (LSTM Based Channel Attention Map) can learn important information by observing the evolution of each sensor channel with time. Notably, it outperforms existing state-of-the-art models, showcasing its superior performance in fetal movement monitoring. We believe that the demonstrated accuracy and efficiency of our model, as outlined in the manuscript, will significantly contribute to advancements in not only in fetal health monitoring but also in introducing a model that brings contextual modifications to robust models that are already in use in computer vision. The integration of novel channel attention module and sensor fusion has aided this introduced model to surpasses current methodologies.

This paper introduces an EEG signal classification approach, leveraging machine learning algorithms. The methodology involves the extraction of features from EEG signal datasets through discrete wavelet transform (DWT). Optimal feature selection is then accomplished using evolutionary algorithms, specifically Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). To identify the most effective classification method, various machine learning algorithms, including Support Vector Machines (SVM), Naive Bayes, Decision Trees, and Random Forest, are systematically compared. This comprehensive evaluation aims to enhance the accuracy and efficiency of EEG signal classification for improved diagnosis and understanding of neurological conditions.

We propose a complete system to enable progressive coding with quality scalability of the mesh geometry, in MPEGâ\euro™s state-of-the-art Video-based Dynamic Mesh Coding (V-DMC) framework. In particular, we propose an alternative method for encoding the subdivision wavelet coefficients in V-DMC, using a mesh-based zerotree coding approach. The proposed method works directly in the native 3D mesh space. It allows us to identify parent-child relationships amongst the wavelet coefficients across different subdivision levels, which can be used to achieve an efficient and versatile coding mechanism. We demonstrate that, given a starting base mesh, a target subdivision surface and a desired maximum number of zerotree passes, our system produces an elegant and visually attractive lossy-to-lossless mesh geometry reconstruction with no further user intervention. Moreover, lossless coefficient encoding with our approach is shown to require almost the same bitrate as the default displacement coding methods in V-DMC. Yet, our approach provides several levels of quality resolution within each target bitrate, while the current solutions encode a single quality level only. To the best of our knowledge, this is the first time that a zerotree-based method has been proposed and demonstrated to work for the compression of dynamic time-varying meshes, and the first time that an embedded quality-scalable approach has been used in the V-DMC framework.

Decoding the directional focus of an attended speaker from listenersâ\euro™ electroencephalogram signals is an important part of a practical brain-computer interface device aimed at improving the quality of life for individuals with hearing impairment. Existing works focus on binary directional focus decoding, i.e., determining whether the attended speaker is on the left or right side of the listener. However, the information brought by the binary decoding to the subsequent speech processing algorithm is limited in practical applications, and more precise decoding of the exact direction of the attended speaker is desired. In this paper, we first present a new dataset with 15 alternative speaker directions, and then demonstrate the feasibility of multiclass directional focus decoding of attended speakers by applying our recently proposed learnable spatial mapping (LSM) module, the benefit of which has already been proven in binary decoding scenarios. Apart from combining the LSM module with the convolutional neural network (CNN), we further validate its benefit by combining it with the spectro-spatial-temporal convolutional recurrent network (CRN), a recently proposed state-of-the-art model for binary directional focus decoding. The proposed LSM-CNN and LSM-CRN models achieve a noteworthy decoding accuracy of 85.7% and 87.5%, respectively, in the presented subject-independent dataset with a decision window length of 1 second. Comprehensive experiments not only substantiate the advantages attributed to the LSM module, but also verify the influence of the decision window length, the distraction caused by interfering speakers, and the contribution of different subband of EEG signals.

The current study presents a novel, non-invasive method for estimating both systolic and diastolic blood pressure by combining photoplethysmogram (PPG) signals with physiological data, such as sex, age, weight, height, heart rate, and BMI, using two Gated Recurrent Units (GRUs) models. The first model processes dynamic patterns in PPG signals, while the second model incorporates physiological parameters. Both models are connected through a series of dense layers. To prepare the datasets for the GRU framework, rigorous preprocessing was conducted. This resulted in a robust architecture capable of accurately predicting systolic and diastolic blood pressure. The proposed method achieved a Mean Absolute Error (MAE) of 1.458 for systolic and 1.164 for diastolic blood pressure. These findings demonstrate the potential of this approach for continual and non-intrusive blood pressure monitoring in wearable health technology. The studyâ\euro™s results also make a significant contribution to the field of medical monitoring technology. The proposed solution addresses a major limitation in traditional blood pressure measurement practices and paves the way for advancements in personalized health monitoring, particularly for managing hypertension and cardiovascular conditions. Â

Objective: In the context of pulmonary rehabilitation exercise training, wearable real-time monitoring of respiratory patterns may represent a valuable tool in increasing accessibility to treatment, as well as expanding the opportunities of treatment automation and locomotor-respiratory coupling. This work explores Hall effect sensing, paired with a permanent magnet, embedded in a chest strap. Methods: Experimental evaluation was performed considering as reference the gold-standard of respiratory monitoring, an airflow transducer, and performance was compared to another wearable device with analogous usability but with a different working principle - a piezoelectric sensor, also embedded in a chest strap. A total of 16 healthy participants performed 15 different activities, representative of pulmonary rehabilitation exercises, simultaneously using the three devices. Evaluation was performed based on detection of flow reversal events, as well as fiducials detection latency. Results: The proposed sensor shows comparable performance to the piezoelectric sensor with a mean ratio, precision, and recall of 1.10, 0.89, and 0.98, respectively, against 1.35, 0.71, and 0.96 of the piezoelectric sensor, overall also presenting consistently smaller latencies. Conclusion: The characterization of the proposed sensor also shows adequate monitoring capabilities for exercises that do not rely heavily on torso mobility, but may present a limitation when it comes to activities such as torso rotations and side stretches. Significance: This work expands the applicability of Hall effect sensors, demonstrating their use in the context of real-time respiratory monitoring.

Our study delves into the challenges of emotion recognition through electroencephalogram (EEG) signals in brain-computer interface systems. Recognizing the limitations of existing methods in accurately capturing intricate emotional patterns in EEG data, we propose a novel approach using asymmetric windowing recurrence plots (AWRP). This technique was designed to enhance the efficiency and accuracy of emotion recognition by encoding EEG signals into detailed image representations that are suitable for advanced deep neural network analysis. Through empirical validations using benchmark datasets (DEAP and SEED), our method demonstrated significant improvements in classification accuracies, notably outperforming existing state-of-the-art methodologies. These findings not only contribute to the field of EEG-based emotion recognition, but also present a novel perspective that can guide future research in neural system analysis and rehabilitation engineering.Â

Construction and analysis of functional brain networks (FBNs) with rs-fMRI is a promising method to diagnose functional brain diseases. Nevertheless, the existing methods suffer from several limitations. First, the functional connectivities (FCs) of the FBN are usually measured by the temporal co-activation level between rs-fMRI time series from regions of interest (ROIs). While enjoying simplicity, the existing approach implicitly assumes simultaneous co-activation of all the ROIs, and models only their synchronous dependencies. However, the functional coactivation is not necessarily always synchronous due to the time lag of information flow and cross-time interactions between ROIs. Therefore, it is desirable to model the asynchronous functional interactions. Second, the traditional methods usually construct FBNs at individual level for feature extraction and classification, leading to large variability and degraded diagnosis accuracy when modeling asynchronous FBN. Third, the FBN construction and analysis are conducted in two independent steps without joint alignment for the target diagnosis task. To address the first limitation, this paper proposes an effective sliding-window-based method to model spatiotemporal FCs in Transformer. Regarding the second limitation, we propose to learn common and individual FBNs adaptively with the common FBN as prior knowledge, thus alleviating the variability and enabling the network to focus on the individual disease-specific asynchronous FCs. To address the third limitation, the common and individual asynchronous FBNs are built and analyzed by an integrated network, enabling end-to-end training and improving the flexibility and discriminativity. The effectiveness of the proposed method is consistently demonstrated on three data sets for MCI diagnosis.

Three-dimensional (3-D) through-the-wall imaging is a challenging topic. It has attracted some research attention in recent years. The 3-D structure is hard to reconstruct because of the limited measurement in the CT-mode imaging method. In order to obtain an accurate 3-D result, the 3-D total variation (3-D TV) algorithm has been adopted. However, the result suffers from image blurring and artifacts. In this paper, a tensor-based optimization framework is proposed to exploit more features of the 3-D wall structure and make up for the shortcomings of the 3-D TV algorithm. The 3-D building structure is modeled as a three-order tensor. Just like the 3-D TV algorithm, the local similarity is considered by the TV regularization constraint to guarantee the reconstruction of the edge. Besides, the group sparsity of the structure is considered to suppress the effect of artifacts and blur. Moreover, in order to keep the global correlation of the image in the case of the errors, the tensor Tucker decomposition is adopted. The performance of this method is discussed in the simulation and real radar data results. It shows that the artifacts and blur are suppressed effectively and the 3-D structure is kept as well.

This paper presents a novel, power- and hardware-efficient, multiuser, multibeam RIS (Reflective Intelligent Surface) architecture for multiuser MIMO, especially suited to operate in very high frequency bands (e.g., high mmWave and sub-THz), where channels are typically sparse in the beamspace and line-of-sight (LOS) is the dominant component. The key module is formed by an active multiantenna feeder (AMAF) with a small number of active antennas, placed in the near field of a RIS with a much larger number of passive controllable reflecting elements. We propose a pragmatic approach to obtain a steerable beam with high gain and very low sidelobes. Then K independently controlled beams can be achieved by closely stacking K such AMAF-RIS modules. Our analysis includes the mutual interference between the modules and the fact that, due to the delay difference of propagation through the AMAF-RIS structure, the resulting channel matrix is frequency selective even in the presence of pure LOS propagation. We consider a 3D geometry and show that “beam focusing” is in fact possible (and much more effective in terms of coverage) also in the far-field, by creating spotbeams with limited footprint both in angle and in range. Our results show that: 1)  simple RF beamforming without computationally expensive baseband digital multiuser precoding is sufficient to practically eliminate multiuser interference when the users are chosen with sufficient angular/range separation, thanks to the extremely low sidelobes of the proposed module; 2) the impact of beam pointing errors with standard deviation as large as 2.5 deg and RIS quantized phase-shifters with quantization bits > 2 is essentially negligible; 3) The proposed architecture is more power efficient and much simpler from a hardware implementation viewpoint than standard RF beamforming active arrays with the same beamforming performance. As a side result, we show also that the array gain of the proposed AMAF-RIS structure grows linearly with the RIS aperture, in line with classical results for standard reflector antennas.

Radar-based human activity recognition (HAR) is a popular research field. Despite claims of high accuracy on self-collected datasets, the ability of these models to handle unexpected scenarios has been largely overlooked. This work introduces a framework for analyzing corruption robustness of radar micro-Doppler spectrogram classification. A set of corruptions are categorized, applied, and systematically tested on common model architectures. Diverse training methods, including adversarial training, cadence velocity diagram (CVD) transformation and data augmentation, are explored. The performance is evaluated on two tasks: indoor HAR and continuous aquatic HAR. Our study unveils several insights. Firstly, relying solely on accuracy may not adequately assess model performance due to dataset limitations. All well-trained models exhibit sensitivity to corruptions. Secondly, deeper convolutional neural network (CNN) models excel in both accuracy and robustness, but confront the problem of overfitting to background. Thirdly, adversarial training enhances robustness against corruptions, albeit at the cost of a slight decrease in accuracy. Lastly, combining data augmentation and adversarial training achieves a balance between accuracy and robustness. In essence, our study contributes to a more profound understanding of the complex interplay between model architecture, classification accuracy, and corruption robustness in radar HAR tasks.

Radar-based human activity recognition (HAR) is a popular research field. In this paper, we explore methods to improve the generalization of micro-Doppler-based swimming activity recognition. We identify three primary challenges for this task: a small dataset, inaccurate period estimation, and an inefficient network design that does not account for the unique characteristics of spectrograms. To address the limited dataset size, we propose spectral data augmentation tailored for micro-Doppler spectrograms. We also investigate two strategies, namely repeated augmentation and contrastive pretraining, to effectively utilize these augmentations. To tackle inaccurate period estimation, we introduce a segmentation approach based on energy distribution to handle temporal period variation, and we include a temporal modeling module in the network structure. To exploit the spread pattern of limb motion in the Doppler dimension and the continuous properties of torso motion in the temporal dimension,  we design a module that consists of both 2D convolution and 1D temporal dynamic convolution to serve as the feature extractor. Our evaluation on a self-collected swimming activity recognition dataset demonstrates that our model achieves high classification accuracy with significantly reduced computational costs. The augmentation methods, particularly when combined with contrastive pretraining, result in improved performance across accuracy and robustness metrics.

A document by Tianyi Wang . Click on the document to view its contents.

Frequency modulation (FM) broadcast signals, as opportunity signals, hold significant potential for indoor and outdoor positioning applications. The existing FM-based positioning methods primarily rely on received signal strength (RSS) for positioning, the accuracy of which needs improvement. In this paper, we introduce an end-to-end FM-based positioning method that leverages deep learning, known as FM-Pnet. This method utilizes the time-frequency representation of FM signals as the network input, allowing the network to automatically learn deep features for positioning. We further propose two strategies, noise injection and enriching training samples, to enhance the model’s generalization performance over long time spans. We construct datasets for both indoor and outdoor scenarios and conduct extensive experiments to validate the performance of our proposed method. Experimental results demonstrate that FM-Pnet significantly outperforms traditional RSS-based positioning methods in terms of both positioning accuracy and stability.

There have been significant efforts devoted to developing advanced mixed-integer linear programming (MILP) solvers, which are powerful tools for solving various real-world optimization problems. Despite the achievements, the limited availability of real-world instances often results in sub-optimal decisions and biased evaluations, which motivates a suite of MILP instance generation techniques. However, these approaches either rely on expert-designed formulations or struggle to capture the rich features of real-world instances. Moreover, the task of generating challenging MILP instances—which are valuable resources for evaluating solvers and motivating more efficient algorithms—remains underexplored. To tackle these problems, we propose G2MILP, which to the best of our knowledge is the first deep generative framework for MILP instances. Specifically, G2MILP represents MILP instances as bipartite graphs and employs a masked variational autoencoder to iteratively corrupt and replace parts of the original graphs to generate new ones. We then propose a hardness-oriented scheme, which iteratively augments the generator by learning from the hardest instances, to enhance G2MILP to construct challenging MILP instances. Experiments demonstrate that G2MILP can generate realistic MILP instances to effectively facilitate downstream tasks. Moreover, G2MILP can generate difficult instances initializing from given datasets, and the boost of hardness can be orders of magnitude.

We integrate the differential signaling technique from OWC systems into our work, suggesting a practical autoencoder-based architecture that accommodates negative encoder output elements. This  capability affords us a greater degree of freedom in shaping the signal constellation space when compared to traditional AEs.

This paper reports a novel design of compact tuneable resonance filter with a highly extinguished and ultra-broad out-of-band rejection in CMOS compatible silicon photonics technology platform. The proposed device is designed with two identically apodized distributed grating structures for guided Fabry-Pérot resonant transmissions in a silicon on insulator rib waveguide structure. The device design parameters are optimized by theoretical simulation for a low insertion loss singly-resonant transmission peak at a desired wavelength.  However, the devices were fabricated (using in-house facilities) to demonstrate multiple resonant transmission peaks along with a singly-resonant one.  We observed that a device length of as low as ∼35 ð?œ‡m exhibits a rejection band as large as ∼60 nm with an extinction of ∼40 dB with respect to the resonant wavelength peak at ð?œ†ð?‘Ÿâˆ¼1550 nm (FWHM ∼80 pm, IL∼2 dB). The experimental results have been shown to be closely matching to our theoretical simulation and modelling results. As expected from the theoretical prediction, the trend pertaining to the trade-off between passive insertion loss and Q-value of the resonances has been observed depending on the device parameters. The thermo-optic tuning characteristics of resonant wavelengths have been obtained by integrating microheaters in the cavity. The resonance peak has been tuned at a rate of 96 pm per mW of consumed thermal power. The thermo-optic switching response has been measured to be in the order of ~5 ð?œ‡s. As a potential application, noise associated with an amplified pump wavelength (ð?œ†ð?‘ƒâˆ¼1550 nm) has been shown to be suppressed by ∼15 dB (up to the detector noise floor) which can be investigated further for large-scale integrated quantum photonic circuits. The demonstrated device can also be explored further for many other applications such as modulation, add-drop multiplexing, sensing etc.Â

Non-autoregressive text-to-speech (TTS) has recently received a lot of atten-tion due to its reliability and fast reasoning. Despite its outstanding achieve-ment, non-autoregressive speech synthesis still faces some critical challenges. A major issue is that non-autoregressive methods necessitate an external toolkit to align the speech with the transcript, thus substantially complicat-ing the process of building the model. Besides, non-autoregressive methods suffer from the one-to-many mapping issue, where the same transcript may correspond to speech in numerous styles. This problem may harm the ex-pressiveness of the generated speech because the model tends to provide output with an average style. To address the above issues, this paper pro-poses a cooperative learning strategy for non-autoregressive speech synthe-sis. Specifically, the suggested method employs both an autoregressive and a non-autoregressive TTS model during the training procedure. The autore-gressive model is trained as a partner at each iteration, providing essential alignment information and also the prosody embedding of the speech to the non-autoregressive model. After receiving the above useful knowledge, the non-autoregressive model can be further trained without relying on external alignment tools. Meanwhile, the prosody embedding from the autoregressive model and the pitch information from the raw audio can be utilised together to alleviate the one-to-many mapping problem. Experimental results demon-strate that our approach can produce comparable speech to the most popular FastSpeech 2 model while drastically reducing the complexity of constructing a non-autoregressive TTS model.

As a machine learning approach, domain adaptation methods are widely applied in cross-scenario fault diagnosis. However, the target domain may need more annotated data, posing challenges to the performance of domain adaptation methods. This paper proposes a fault diagnosis method based on hybrid physics and domain adaptation (HPDA) with its application to marine current turbines (MCTs) scenarios. Specifically, this method first establishes a rotational feature alignment model based on physical variables. Then, it aligns the feature of the target domain data with physical parameters. Finally, an augmented domain adversarial model is trained using pre-alignment samples. Data from MCT prototypes are collected to validate the effectiveness of the proposed method. Experimental results demonstrate the proposed method’s superior stability and data transferability compared with the state-of-the-art.