IEEE Transactions on Biomedical Engineering最新文献

Objective: Cardiovascular diseases (CVDs) are a leading global health threat. The automatic classification of phonocardiogram (PCG) signals is crucial for their early diagnosis, yet existing models are often limited by analyzing features from only a single domain (time or frequency), failing to fuse complementary information. This study aims to develop a model that overcomes this limitation by effectively integrating both time-domain and frequency-domain features to improve classification accuracy and robustness.

Methods: We propose a novel end-to-end dual-branch deep learning model. The time-domain branch utilizes a 1D Convolutional Neural Network (CNN) with Transformer blocks to capture instantaneous dynamics and long-range dependencies. The frequency-domain branch uses a ResNet to extract robust spectral patterns from Mel-spectrograms. A key innovation is our bidirectional cross-attention fusion module, which facilitates deep interaction and mutual enhancement between the two feature modalities. Furthermore, we employ a transfer learning strategy to ensure robust performance on smaller or more challenging datasets.

Results: Comprehensive evaluations on multiple public datasets demonstrate that our model achieves state-of-the-art (SOTA) performance. On the 2016 PhysioNet Challenge dataset, it reached an accuracy of 98.86% and an F1-score of 97.19%, significantly outperforming existing baseline methods.

Conclusion and significance: Our dual-branch fusion model provides a more effective and robust framework for heart sound classification. This work offers strong support for the development of highly accurate automated tools for the auxiliary diagnosis of CVDs, thereby holding the potential to enhance early detection and improve clinical outcomes.

Subject-specific finite-element analysis (FEA) models enable accurate simulation of vertebral biomechanics but are often time-consuming to construct and solve under varying conditions. This study presents a novel deep learning (DL)/machine learning (ML)-based surrogate model that predicts stress distributions in vertebral bodies with high efficiency. The model integrates vertebral shape encoding and employs separate decoding branches for surface and internal nodes. It was trained on 3,960 synthetic L1 vertebrae generated via data augmentation from 42 real computed tomography (CT) scans. Evaluation on independent test samples yielded a mean absolute error (MAE) of 0.0596 MPa and an R² of 0.864 for von Mises stress. Visualization results confirm strong agreement between predicted and FEA-computed stress patterns, with localized discrepancies observed at the anteroinferior margin and pedicles. Moreover, an end-to-end automated pipeline was established based on the developed model, reducing the total processing time from 90-120 min to approximately 134-154 s per subject. These findings highlight the potential of the proposed surrogate model to facilitate rapid, subject-specific biomechanical assessments in clinical workflows.

Objective: This study aims to develop a flexible, implantable neural probe with tunable stiffness and multifunctionality for electrophysiology, drug delivery, and optogenetics, while minimizing immune response.

Methods: The probe was fabricated from polydimethylsiloxane (PDMS) with a U-turn polyester-filled microchannel (30 μm × 30 μm, 14.7 mm length). Polyester provides rigidity at room temperature and softens near body temperature for tissue compatibility. Mechanical simulations optimized probe dimensions (60 μm × 300 μm × 7 mm), ensuring insertion forces above 1.5 mN. A microfluidic mixer (90 μm × 30 μm, 7 mm) was integrated for controlled drug delivery. Heat transfer and fluid simulations assessed thermal stability and laminar flow. The device also included four gold electrodes, a bio-amplifier, and a blue μ-LED.

Results: Experiments confirmed stable channel performance with no leaks or bubbles, effective heat management at 37°C, and a 60° tip angle as optimal for insertion. The probe consumed ∼46.6 mW and enabled reliable electrophysiological recordings and fluorescence activation in vitro. Biocompatibility testing validated its suitability for long-term implantation.

Conclusion: The PDMS-polyester probe achieves thermally controlled stiffness, precise drug delivery, and integrated optogenetic stimulation while maintaining stable electrophysiological recording performance.

Significance: This work introduces a multifunctional neural probe that addresses limitations of rigid implants by combining flexibility, drug delivery, and optogenetics. The platform has strong potential for advancing long-term neuroengineering applications, reducing tissue response, and enabling multimodal brain interfacing.

Objective: The blood-brain barrier (BBB) poses a significant challenge for central nervous system drug delivery due to its selective permeability. Focused ultrasound (FUS) combined with bubble agents enables non-invasive, targeted BBB disruption. However, current methods for assessing efficacy and safety have limitations of high cost, low temporal resolution, or moderate predictive reliability, etc. Methods: This study proposes a novel gated attention-based model (GAB) for predicting BBB opening outcomes using time-domain acoustic signal clips. The GAB architecture incorporates an acoustic encoder to extract spectral and temporal features from short clips while modeling dependency between adjacent clips. A multi-way gated attention mechanism aggregates clip-level features, enhancing inter-class discriminability through adaptive selection. A task-specific loss function combining classification and similarity constraints further improves prediction by reducing redundancy in attention patterns. The outcomes of 174 FUS treatments were classified into three categories: BBB not opened, BBB opened without hemorrhage, and BBB opened with hemorrhage.

Results: The GAB achieved superior performance in hemorrhage prediction (accuracy = 86.6 ± 4.9%, recall = 82.1 ± 10.9%, AUC = 0.894 ± 0.036, F1 score = 0.846 ± 0.062) compared to traditional cavitation dose-based methods and our previously developed frequency-domain deep learning models. Visualization of attention weights revealed that the model effectively distinguished broadband inertial cavitation signals (associated with hemorrhage) from ultra-harmonic stable cavitation signals (linked to safe BBB opening).

Conclusion: The proposed method enhances temporal resolution, achieves superior predictive performance with interpretable cavitation features, and shows strong potential for real-time outcome prediction in FUS-mediated BBB disruption.

Robotic ultrasound imaging systems primarily focus on enhancing automation in grayscale image acquisition but lack essential functional information, which restricts their clinical effectiveness and efficiency. In this regard, we propose a novel robotic ultrasound system capable of automatically screening and anomaly localization using quasi-static elastography (QE). For continuous screening, a compliant force control strategy is devised to manage the complex probe operations required for elasticity data acquisition. This involves the integration of adaptive out-of-plane posture control and in-plane palpation motion control. For anomaly localization, tissue strains are analyzed using multi-source motion data. We introduce an unsupervised tissue displacement estimation method, complemented by a strain estimator with multi-frame fusion for robust strain estimation. A 3D strain map is reconstructed to enable closed-loop control for automated robotic localization. The system has been validated through extensive experiments on two realistic phantoms and tested on human subjects. Results demonstrate that our system can perform robotic QE-based screening across subjects with varying surface conditions and lesion depths, showing improved efficiency and adaptability compared to existing systems. It achieves satisfactory accuracy in strain-based anomaly localization, with a detection rate of 0.77 and an average localization error of 1.01±0.47 mm for the abdominal phantom, and 0.73 and 3.44±0.84 mm for the more challenging thyroid phantom. By identifying and localizing suspicious anomalies in 3D space, the proposed system shows promise in providing preliminary dia.

Objective: This study proposes an Improved Deep Transfer Network (IDTN) to enhance decoding accuracy, calibration efficiency, and adaptability of intracortical brain machine interface (iBMI) systems while reducing the reliance on new labeled samples.

Methods: IDTN integrates two core components: Structural Joint Discriminative Maximum Mean Discrepancy (SJDMMD) and Kernel Norm Improved Multi-Gaussian Kernel (KNK). SJDMMD extends the standard MMD framework by incorporating a structure-enhanced soft label weighting mechanism that simultaneously minimizes intra-class distributional shifts and maximizes inter-class margins for precise cross-domain alignment. KNK employs multi-Gaussian kernels with kernel norm regularization to enhance high-dimensional feature representations and sharpen inter-class boundaries, thereby improving the effectiveness of SJDMMD.

Results: Evaluated on neural datasets from two rhesus macaques, IDTN achieved superior performance in both intra- subject and inter-subject transfer scenarios, consistently outperforming state-of-the-art methods in decoding accuracy. IDTN also exhibited consistent decoding stability across daily recording sessions. Ablation studies further confirm that SJDMMD improves inter-class separability and intra-class coherence, while KNK contributes to more effective kernel mapping in complex feature spaces.

Conclusion: These findings underscore the effectiveness of structure-aware transfer learning for neural decoding.

Significance: They also highlight the potential of IDTN for deployment in real-world iBMI applications, particularly in data-limited or cross-subject environments.

Current learning-based image denoising methods have achieved impressive performance. However, their reliance on deep neural architectures and large paired datasets limits their applicability in data-limited or edge computing scenarios. Motivated by the expressive functional approximation power of Kolmogorov-Arnold networks (KANs), here we present ZS-KAN-a lightweight yet highly effective and computationally efficient zero-shot denoising method. ZS-KAN combines the computational efficiency of convolutional neural networks with the representational flexibility of KANs, achieving competitive denoising performance while requiring only 1%-25% of the parameters used by other recent zero-shot approaches. Experimental results on synthetic and real-world noisy data demonstrate that ZS-KAN achieves comparable or even superior performance to state-of-the-art zero-shot methods while maintaining significantly lower model complexity. These advantages highlight the potential of ZS-KAN for practical deployment. The PyTorch implementation is publicly available at: https://github.com/Jayx-Wang/ZS-KAN.

The increasing availability of large electroencephalography (EEG) datasets enhances the potential clinical utility of deep learning (DL) for cognitive and pathological decoding. However, dataset shifts due to variations in the population and acquisition hardware can considerably degrade the model performance. We systematically investigated the generalisation of DL models to unseen datasets with different characteristics, using age as the target variable. Five datasets were used in two different experimental setups, including (1) leave-one-dataset-out (LODO) and (2) leave-one-dataset-in (LODI) cross validation. A comprehensive set of 1805 different hyperparameter configurations was tested, including variations in the DL architectures and data pre-processing. The performance varied across source/target dataset pair. Using LODO, we obtained Pearson's r values of {0.63, 0.84, 0.75, 0.23, 0.10} and $R^{2}$ values of {-0.01, 0.63, 0.41, -4.66, -70.98}. For LODI, the results varied in Pearson's r from -0.11 to 0.84 and $R^{2}$ values from -704.89 to 0.65, depending on the source and target dataset. Adjusting the model intercepts using the average age of the target dataset substantially improved some $R^{2}$ scores. Our results show that DL models can learn age-related EEG patterns which generalise with strong correlations to datasets with broad age spans. The most important hyperparameter was to use the frequency range between 1 and 45Hz, rather than a single frequency band. The second most important hyperparameter effect depended on the experimental setup. Our findings highlight the challenges of dataset shifts in EEG-based DL models and establish a benchmark for future studies aiming to improve the robustness of DL models across diverse datasets.

Objective: Global Maxwell Tomography (GMT) is a noninvasive inverse optimization method for the estimation of electrical properties (EP) from magnetic resonance (MR) measurements. GMT uses the volume integral equation (VIE) in the forward problem and assumes that the sample has negligible effect on the coil currents. Consequently, GMT calculates the coil's incident fields with an initial EP distribution and keeps them constant for all optimization iterations. This can lead to erroneous reconstructions. This work introduces a novel version of GMT that replaces VIE with the volume-surface integral equation (VSIE), which recalculates the coil currents at every iteration based on updated EP estimates before computing the associated fields.

Methods: We simulated an 8-channel transceiver coil array for 7 T brain imaging and reconstructed the EP of a realistic head model using VSIE-based GMT. We built the coil, collected experimental MR measurements, and reconstructed EP of a two-compartment phantom.

Results: In simulations, VSIE-based GMT outperformed VIE-based GMT by at least 12% for both EP. In experiments, the relative difference with respect to probe-measured EP values in the inner (outer) compartment was 13% (26%) and 17% (33%) for the permittivity and conductivity, respectively.

Conclusion: The use of VSIE over VIE enhances GMT's performance by accounting for the effect of the EP on the coil currents.

Significance: VSIE-based GMT does not rely on an initial EP estimate, rendering it more suitable for experimental reconstructions compared to the VIE-based GMT.

Objective: Digital subtraction angiography (DSA) is the gold standard modality for diagnostics and guidance for interventional procedures. Spectral imaging has previously been explored for DSA, but severe noise amplification from material decomposition has impeded clinical adoption. We present a novel joint processing strategy that leverages both temporal and spectral information for material decomposition to address this issue.

Methods: We develop a model-based material decomposition approach that utilizes the pre- and post-contrast images simultaneously for material estimation. Performance was evaluated on a small-vessel phantom on a test bench with a photon-counting detector. Joint processing was compared with temporal subtraction and previously proposed spectral DSA techniques including hybrid subtraction and conventional three-material decomposition. Additional simulation was performed to investigate performance with perfectly calibrated spectral response and sensitivity to patient motion.

Results: The improved conditioning of the proposed method effectively reduces bias and noise in the spectral results and allows three-material decomposition with dual-energy spectral measurements. The method achieved more than an order of magnitude variance reduction compared to previously proposed spectral DSA techniques. Compared to temporal subtraction, a mean variance reduction of 23.9% was achieved in simulation and 10.8% in experimental data. The degree of reduction is object-dependent. Noise reduction achieved in physical experiments is slightly lower than that in simulation, likely due to bias from imperfect spectral calibration. The method is equally sensitive to motion compared to temporal subtraction.

Conclusion: The proposed method addresses a major image quality challenge limiting previous approaches and outperforms temporal subtraction.

Significance: Such improvements facilitate the clinical translation of spectral angiography.