IntroductionThis study proposes advanced neural network architectures for classifying specific motor-related electroencephalography (EEG) tasks, employing deep feature extraction techniques. We analyzed EEG data from the MILimbEEG dataset, consisting of recordings from 60 individuals as they performed eight distinct motor movements: baseline with eyes open, left-hand closing, right-hand closing, dorsiflexion and plantarflexion of both the left and right feet, as well as rest periods between tasks. The high precision achieved in this study underscores the efficacy of sophisticated computational models like the GMDH network in enhancing the interpretation of EEG signals for the development of brain-computer interfaces (BCIs). This research significantly advances the potential of EEG as a reliable modality for BCIs, effectively translating brain activity into actionable commands suitable for neurorehabilitation and assistive technologies.MethodsFor each of the 16 electrodes used in the recordings, 10 critical features were extracted, resulting in a comprehensive set of 160 features per sample that encapsulate the intricate brain activities associated with each task. A Group Method of Data Handling (GMDH) neural network, structured with eight hidden layers and a decremental arrangement of neurons from 40 in the first to 5 in the last, was utilized to classify these tasks.ResultsThis network configuration achieved an impressive classification accuracy of approximately 96%, demonstrating a robust capability to accurately decode EEG signals tied to specific motor actions.DiscussionThe high precision achieved in this study underscores the efficacy of sophisticated computational models like the GMDH network in enhancing the interpretation of EEG signals for the development of brain-computer interfaces (BCIs). This research significantly advances the potential of EEG as a reliable modality for BCIs, effectively translating brain activity into actionable commands suitable for neurorehabilitation and assistive technologies. Our findings contribute substantially to the BCI field, promising to improve clinical outcomes by enabling more precise and effective interaction with neurorehabilitation devices.

Sewer defect recognition is a critical foundation for urban drainage systems, by analyzing the video in the sewer to find the problems. Contrastive Language-Image Pre-training model(CLIP) performs well on general vision tasks but misses the fine-grained structural variations and localized defect features, resulting in limited performance in practical sewer defect classification. Therefore, a CLIP based multi-label sewer defect classification method is proposed, which leverages the transfer capability of large language model and integrates fine-grained visual-linguistic features. To tackle the problem of insufficient fine-grained defect feature extraction, the Prompt-based Contextual Representation Construction (PCRC) module is designed, leveraging learnable prompts and a two-stage modeling strategy to capture fine-to-coarse contextual representations for each category. Furthermore, the Feature-Level Matching (FLM) module is introduced to align the fine-grained image-text feature for improving defect recognition accuracy. Finally, the ablation studies and extensive comparisons with advanced methods on the public dataset Sewer-ML is presented. Experimental results demonstrate that the proposed approach achieves state-of-the-art performance, of which the mAP and F1-score achieve 75.02% and 80.08%, respectively.

ECG is an important signal for cardiovascular disease prediction. Since the ECG signals are often stored as images in clinical practice, we transformed sequential ECG data into images and evaluated the performance of single-modal and multi-modal fusion methods with ECG and EHR data. Results indicate that EHR features are more informative than ECG images, and multimodal fusion method outperforms all single-modal methods.

This paper presents a high-performance, resource-efficient digital implementation of the Dressed Neuron Model (DNM), a biologically inspired system that captures bidirectional interactions between neurons and astrocytes. Unlike classical neuron models, the DNM incorporates astrocyte-mediated calcium and IP3 signaling, forming a closed-loop feedback system capable of exhibiting spontaneous, seizure-like oscillations. To address the high computational complexity of this model on hardware, we introduce a Hybrid Model of DNM (HMoDNM), which approximates all major nonlinearities using a combination of dual-sinusoidal expressions and ROM-based lookup tables. These approximations achieve high numerical fidelity with root mean square error (RMSE) below 10-2 for most functions, while ensuring hardware efficiency. The full system is implemented on a Xilinx Zynq-7000 XC7Z010 FPGA using a shift-and-add architecture with zero DSP slice utilization. All signals are represented in a fixed-point format < 1,10, 27 >, with dynamic range coverage up to 800 mu M for IP3 and 600 mu M for calcium. The design includes pipelined neuron and astrocyte cores, clock-gated nonlinear units, and shared computation modules, achieving a maximum clock frequency of 305 MHz and throughput of 21.7 million Euler steps per second. Overall resource usage is 6690 LUTs (38.0%), 2640 FFs (7.5%), and 4 BRAMs (6.7%), with a low dynamic power consumption of 167 mW and operating temperature of 35.1 degrees C at room ambient. To validate the model's functional accuracy, we compare the HMoDNM outputs against the original DNM across two dynamic regimes, achieving correlation coefficients above 94% and NRMSE values below 0.06 for membrane voltage, calcium, and IP3. Designed specifically for epilepsy modeling, this architecture provides a robust foundation for real-time tracking and control of astrocyte-influenced seizure dynamics. The proposed HMoDNM architecture offers a versatile foundation for hardware-based neuromorphic applications, including real-time seizure detection, closed-loop neurostimulation systems, and low-power embedded platforms for modeling neuron-glia interactions in brain-inspired computing.

As a State Space Model (SSM) that achieves long-range dependency modeling with linear computational complexity, Mamba demonstrates significant efficiency advantages in medical image segmentation. However, while Mamba-based methods enable long-range modeling with linear complexity, their global dependency mechanisms often lead to local feature attenuation, particularly affecting the processing of complex anatomical structures. Existing multi-scale fusion methods also exhibit limited compatibility with State Space Models. To address these challenges, this paper proposes the Local-Global Fusion Vision Mamba UNet (LGFVM-UNet) framework. Its core innovation lies in the Dynamic Gating-enhanced Local-Global Fusion Visual State Space (LGF-VSS) block, which enables the synergistic modeling of global context and local details. Additionally, we designed a Multi-level Cross-scale Feature Fusion Block (MCFB) that enhances multi-scale feature representation through bidirectional resampling and spatial-channel dual attention mechanisms. Additionally, we propose a Gradient Statistics-based Adaptive Hierarchical Loss that dynamically adjusts multi-level supervision weights to optimize the learning process. The proposed method is experimentally validated on five public medical image segmentation datasets spanning diverse imaging modalities and anatomical structures. Results demonstrate that our approach outperforms state-of-the-art methods, excelling in long-range dependency modeling, local detail capture, and multi-scale feature fusion. The source code of our work is available at https://github.com/NicoleDyson/LGFVM-UNet.

In real clinical settings, medical image datasets are often partially annotated due to high labeling costs and complexity, which limits multi-label classification. Existing methods often attempt to tackle this challenge by either decoupling features to generate pseudo-labels or by treating all unknown labels as negative labels during training. In scenarios with severe label scarcity, the former approach may fail to generate high-quality decoupled features, while the latter is prone to introducing label noise. To address these challenges, we propose a novel method for partial multi-label medical image recognition tasks, called Asymmetric Dual Thresholds and Co-occurrence Relationship (ADTCR). Specifically, ADTCR consists of two pseudo-label generation strategies: Asymmetric Dual Threshold (ADT) and Co-occurrence Relationship (CR). The ADT strategy is designed to initially identify pseudo-labels by applying a lower threshold for negative pseudo labels and a higher threshold for positive pseudo labels, ensuring the generation of high-quality pseudo labels. Meanwhile, the CR strategy aims to uncover potential positive labels by capturing label co-occurrence relationships, enabling the detection of latent positive labels among the unknown ones. Finally, to assess the model's confidence in the generated pseudo-labels, we design a Threshold-based Weighting Loss (TWL), which uses threshold-based weights to weight each pseudo-label, thereby further improving performance. Extensive experiments conducted on three multi-label medical image datasets, i.e., Axial Spondyloarthritis, NIH Chest X-ray 14, ODIR-5K, demonstrate that our method achieves state-of-the-art performance.

This paper introduces a nonlinear, bio-inspired cochlear network based on a two-dimensional array of coupled Hopf resonators, developed and implemented for real-time cochlear implant (CI) applications. The proposed model captures the key phenomena of the human cochlea — active amplification, nonlinear compression, and traveling-wave propagation — while maintaining a compact and computationally efficient structure suitable for hardware realization. Each resonator represents a local frequency channel coupled through diffusive interactions, forming a network that reproduces tonotopic mapping, synchronization, and frequency selectivity consistent with physiological data. Comprehensive simulations demonstrate stable spatio-temporal dynamics, robust traveling-wave behavior, and high noise tolerance with signal-to-noise ratios exceeding 26dB and correlation coefficients above 0.95 under additive and colored noise perturbations. The model’s frequency-domain analysis further enables bio-inspired mapping of cochlear responses into multichannel stimulation envelopes for CI electrodes. A real-time FPGA implementation validates the framework, achieving low-latency operation and efficient resource utilization (less than 45% of logic and DSP resources) while reproducing electrode outputs with under 1.5% deviation from simulation. This work bridges nonlinear cochlear dynamics and neuromorphic hardware, providing a biologically faithful and hardware-efficient foundation for next-generation cochlear implant signal processors.

The cyber–physical integration of modern electrical power and energy systems increasingly demands high-performance, real-time computational tools. Bio-inspired computing paradigms, such as Spiking Neural Networks (SNNs), offer a highly energy-efficient framework for handling complex, data-driven analytics in these systems. However, their hardware implementation remains a significant challenge, heavily constrained by computational speed and power consumption. The Izhikevich model, celebrated for its rich nonlinear dynamics, exemplifies this issue; its resource-intensive quadratic term typically demands costly hardware multipliers, limiting scalability in power-constrained environments. This paper presents a highly optimized, low-power, and multiplier-free digital architecture for synaptic-coupled Izhikevich models on FPGAs, tailored to address the computational efficiency required by next-generation cyber–physical systems. Instead of a direct computation, we reformulate the model’s quadratic dynamics into a trigonometric representation. This transformation enables the deployment of a high-speed, customized CORDIC engine, effectively replacing the expensive multipliers. This refined geometric approach not only preserves the full spectrum of dynamic behaviors but also yields substantial performance gains. Our FPGA implementation demonstrates a significant 4.24-fold speedup and an 8% reduction in power consumption when benchmarked against conventional designs, while also showing notable improvements over other CORDIC-based approaches. This work provides a scalable and highly energy-efficient hardware pathway for developing real-time computational engines capable of complex, large-scale simulations in modern energy infrastructures.

Cardiovascular disease (CVD) is major global health concern. Their complex etiology makes accurate detection particularly challenging. Multimodal data can offer complementary insights from different perspectives, thereby enhancing detection accuracy. However, existing multimodal detection methods overlook the inherent differences between modalities, leading to suboptimal fusion effects. Inspired by visual-language model (VLM), we propose a fusion network based on Semantic-guided Modal Alignment (SMA), utilizing disease diagnosis labels as prompts to align electrocardiogram features with electronic health record features. This approach enables consistent high-level semantic representation across the two modalities, thereby enhancing the effectiveness of multimodal fusion. The experimental results on MIMIC-IV-ECG myocardial infarction (MI) dataset indicate that our model outperforms baseline models with an accuracy of 89.81% and an AUC of 91.24%.

Axial spondyloarthritis (axSpA) is a chronic inflammatory disease affecting the sacroiliac joints and spine, and delayed diagnosis can lead to irreversible structural damage and functional impairment. In practice, multimodal modeling remains challenging due to subtle cross-slice lesion patterns in SIJ-MRI, redundant modality-specific features, and patient-level variation in modality reliability. To address these issues, we propose ACSD-Net, a multimodal framework for ASAS-guided axSpA auxiliary diagnosis. The model contains three parts: a Slice Sequence Fusion Module (SSFM) for cross-slice SIJ-MRI encoding, a Sparse Feature Filtering Module (SFFM) for refining modality-specific representations, and a Confidence-Driven Fusion Module (CDFM) for sample-wise multimodal fusion. We evaluated ACSD-Net on a multicenter dataset of 466 subjects from three institutions using T1-weighted SIJ-MRI, T2-weighted SIJ-MRI, and six clinical variables. In five-fold cross-validation, ACSD-Net achieved 84.13% accuracy, 90.42% AUC, 89.13% F1-score, and an MCC of 0.6026. Under a fixed unseen-center evaluation protocol, it achieved 82.89% accuracy, 90.94% AUC, and an MCC of 0.6002. These results show that ACSD-Net can integrate imaging and clinical information effectively for multimodal axSpA diagnosis.

This paper proposes an innovative approach to translating the nonlinear dynamics of a memristive FitzHugh-Nagumo-Hindmarsh-Rose (FN-HR) coupled neuron model into an AI-optimized, resource-efficient VLSI implementation on FPGA platforms, advancing intelligent computing paradigms. The bidirectional memristive synapse coupling FN and HR neurons enables rich dynamic behaviors such as mixed-mode oscillations and chaos, which are harnessed to enhance adaptive machine learning and neural network training. A detailed dynamical analysis, including Lyapunov exponent spectra and synchronization properties, identifies parameter regimes suitable for AI applications. Nonlinear operators are approximated using quantized lookup tables and three-term sinusoidal expansions, achieving RMSE values of 0.0105 (FN) and 0.0114 (HR) while eliminating DSP usage. Synthesized on an AMD Zynq UltraScale+ ZCU104 FPGA, a 50-neuron network utilizes 5.3% LUTs and 7% BRAM, delivering 42 million neuron-updates per second at 210 mW. This work establishes a scalable, low-power platform for real-time AI-driven neuromorphic computing and intelligent adaptive control systems.

The primary objective of this study is to investigate information transmission patterns within the motor control system across varying grip strength levels and to address analytical challenges arising from false connections and inherent complexity in cortico-muscular functional networks. We propose the OPTN-HHKD method as a novel framework for cortico-muscular coupling (CMC) analysis. The ordinal partition transition networks (OPTN) method was employed to construct effective connectivity networks and mitigation of false connection. Furthermore, Helmholtz-Hodge-Kodaira decomposition (HHKD) was applied to decompose the networks into circular and gradient flows to highlight important channels in information transmission. Innovatively, causal hierarchy was integrated into the CMC analysis to intuitively reflect causal relationships between the cortex and muscles in the motor control system. The effective connectivity network constructed by OPTN showed task-specificity and frequency-specificity of CMC. Gradient flow networks revealed that information predominantly converged into the flexor digitorum superficialis and the auxiliary muscles related to the grasping task, while in the γ-band, information was generally transmitted from the muscles to the motor cortex. The causal hierarchy intuitively revealed descending coupling associated with β rhythm and ascending coupling linked to γ rhythm. The proposed framework effectively addresses current limitations in CMC analysis, offering a novel perspective for studying human motor control mechanisms.

IntroductionUnderstanding the neural mechanisms underlying general anesthesia remains a significant challenge in neuroscience and clinical practice. Traditional software-based simulations of large-scale brain networks are often constrained by high computational costs and fail to achieve real-time performance.MethodsIn this paper, we propose a high-performance hardware implementation of large-scale neuromorphic system to investigate anesthetic-induced neural dynamics. The system successfully models a cortical network comprising 10,000 spiking neurons (8,000 excitatory and 2,000 inhibitory) utilizing the biologically plausible Izhikevich neuron model. Deployed on a field-programmable gate array (FPGA), the proposed architecture exploits high parallelism to achieve real-time simulation speeds. By adjusting synaptic weights and network parameters to mimic the pharmacological eects of anesthetic agents, our system can continuously monitor and evaluate state transitions in neural synchronization and firing patterns.ResultsThe results demonstrate that the hardware-accelerated neuromorphic approach provides an efficient, scalable, and real-time platform for investigating large-scale neural dynamics.DiscussionPending future validation against empirical clinical EEG data, this foundational framework paves the way for advanced brain–machine interfaces and closed-loop anesthetic delivery systems.

Lightweight medical image segmentation models are essential for real-world clinical deployment, where computational resources and latency are strictly constrained. However, existing compact architectures often suffer from insufficient feature representation, weak contextual modeling, and performance degradation when handling complex anatomical structures. To address these limitations, we propose DHR-Net, an ultra-lightweight segmentation network designed to achieve high accuracy while maintaining extremely low computational cost. The framework integrates three complementary modules: DSDC, a dynamic staged depthwise convolution block that enhances multi-scale local feature extraction with minimal GFLOPs; HA, a hybrid attention mechanism that adaptively separates and refines relevant and irrelevant channel groups; ResECA, a lightweight cross-channel semantic enhancement module that preserves global context at the encoder-decoder bottleneck without dimensionality reduction. We conducted comprehensive experiments on five publicly available medical datasets (Kvasir-Seg, CVC-ClinicDB, DDTI, ISIC2018, and Synapse). Compared with UNext, DHR-Net achieves superior segmentation performance while requiring fewer parameters and lower computational complexity. Notably, DHR-Net is among the very few ultra-lightweight models that maintain both the parameter count and computational cost at only 0.06M parameters and 0.06 GFLOPs, respectively. Ablation studies further verify the effectiveness and complementarity of the DSDC, HA, and ResECA modules. Owing to its efficiency, modularity, and strong generalization capability, DHR-Net provides a practical and deployable solution for clinical scenarios such as point-of-care devices, real-time analysis, and resource-limited medical environments. Code at https://github.com/Phil-y/DHR-Net.

This paper presents a high-performance, resource-efficient digital implementation of the Dressed Neuron Model (DNM), a biologically inspired system that captures bidirectional interactions between neurons and astrocytes. Unlike classical neuron models, the DNM incorporates astrocyte-mediated calcium and IP3 signaling, forming a closed-loop feedback system capable of exhibiting spontaneous, seizure-like oscillations. To address the high computational complexity of this model on hardware, we introduce a Hybrid Model of DNM (HMoDNM), which approximates all major nonlinearities using a combination of dual-sinusoidal expressions and ROM-based lookup tables. These approximations achieve high numerical fidelity with root mean square error (RMSE) below 10−2 for most functions, while ensuring hardware efficiency. The full system is implemented on a Xilinx Zynq-7000 XC7Z010 FPGA using a shift-and-add architecture with zero DSP slice utilization. All signals are represented in a fixed-point format 〈1,10,27〉, with dynamic range coverage up to 800 μM for IP3 and 600 μM for calcium. The design includes pipelined neuron and astrocyte cores, clock-gated nonlinear units, and shared computation modules, achieving a maximum clock frequency of 305 MHz and throughput of 21.7 million Euler steps per second. Overall resource usage is 6690 LUTs (38.0%), 2640 FFs (7.5%), and 4 BRAMs (6.7%), with a low dynamic power consumption of 167 mW and operating temperature of 35.1 °C at room ambient. To validate the model’s functional accuracy, we compare the HMoDNM outputs against the original DNM across two dynamic regimes, achieving correlation coefficients above 94% and NRMSE values below 0.06 for membrane voltage, calcium, and IP3. Designed specifically for epilepsy modeling, this architecture provides a robust foundation for real-time tracking and control of astrocyte-influenced seizure dynamics. The proposed HMoDNM architecture offers a versatile foundation for hardware-based neuromorphic applications, including real-time seizure detection, closed-loop neurostimulation systems, and low-power embedded platforms for modeling neuron–glia interactions in brain-inspired computing.

Most RGB-D-T SOD models are dedicated to establishing complex feature extraction and multi-modal fusion modules to tackle scenarios that RGB SOD models can't solve, such as insufficient lighting, complex background, and ignoring the portability and mobility of RGB-D-T SOD models. We propose a lightweight, multi-modal salient object detection model, LMFNet, to facilitate practical applications. We reduce the base feature channels and focus on building a multi-level, multi-scale feature fusion module to ensure that the model can effectively extract the most discriminative features across multiple scales, achieving overall low-parameter and high-efficiency performance. Concretely, we use three parameter-sharing encoders to extract features from the RGB, depth, and thermal modalities, which significantly reduces the number of model parameters required. In addition, to simplify the multi-modal fusion task, we designed an efficient fusion module that achieves feature integration with minimal computational overhead by concatenating features followed by depth-wise separable convolution. The fusion process follows a sequential order: first, thermal and depth characteristics are fused, then the fused result is combined with RGB characteristics. This approach helps reduce the risk that important RGB information is overshadowed during the fusion process. Then, to improve the model's performance, we fuse features from different levels at varying scales, allowing each layer of features to contain rich detail and semantic information. Different levels incorporate features at different scales. Finally, we employ a decoder to progressively decode the multi-level features obtained in the previous stage. In this process, the multi-scale fused features are integrated from coarse to fine, gradually predicting precise salient object boundaries. Extensive experiments demonstrate that our approach strikes an optimal balance between processing speed and performance. Specifically, our model, with only 2.7M parameters, achieves a processing speed of 202 FPS on an NVIDIA 2080Ti with a 352 x 352 input size while maintaining superior performance. The code is available at https://github.com/banjamn/LMFNet.

Current analyses of intermuscular coupling (IMC) predominantly rely on pairwise metrics, which may fail to capture high-order interactions (HOIs) essential for understanding complex neuromuscular coordination. This study develops a multi-domain high-order analysis framework for IMC by integrating O-information rate (OIR), frequency-domain O-information rate (fOIR), and $\beta $ -band B-index-rate network analysis to characterize high-order dependencies among muscles across the time domain, frequency domain, and connectivity structure, with a focus on activation and coordination patterns during natural grasping/lifting tasks. The results reveal substantial information redundancy within proximal muscle groups, and synergistic interactions typically emerged from combinations of muscles across different groups. Task- and posture-dependent variations in high-order dependency patterns were identified, with stronger overall muscle redundancy observed during lifting tasks than during gripping tasks. Frequency-domain analysis showed that high-order dependencies were most consistently expressed in the 15-30 Hz range ( $\beta $ band), where redundancy/synergy effects were evident during upper-limb grasping/lifting tasks. The proposed framework to a certain extent reveals task-dependent high-order interactions under partially controlled biomechanical conditions, highlighting redundancy within several muscle groups and synergy across groups. These findings advance the understanding of complex neuromuscular interactions and provide insights for motor control systems and rehabilitation.

This study proposes a smart energy system for dynamically supplying electrical and thermal demands of a 1,400-bed tertiary-care hospital using an integrated 2.5 MW geothermal–transcritical CO2 Rankine combined heat and power (CHP) configuration coupled with supervised machine learning. A validated thermodynamic model of a double-flash geothermal cycle integrated with a transcritical CO2 Rankine subsystem was developed and benchmarked against published data, yielding deviations below 0.1% in net power and efficiency. Parametric analysis revealed an optimal pressure ratio range of 1.6–1.7, where net power reaches approximately 2,580 kW, while total heating varies between 11,940 and 12,360 kW depending on operating conditions. A polynomial regression surrogate model (R2 = 99.79% for power and 100% for heating) was constructed using 100 simulation points to enable rapid inverse optimization. The smart control framework dynamically adjusts pressure and pressure ratio to match hourly hospital loads across three operational shifts. Results show that the system consistently satisfies electrical demands (up to 20,480 kW per shift) and heating requirements (up to 99,550 kW per shift), while generating a daily surplus of 1,979 kW electricity and 3,030 kW heating. The proposed supervised machine learning–assisted framework demonstrates high predictive accuracy, operational flexibility, and suitability for resilient healthcare energy infrastructure.

This paper proposes two efficient lightweight segmentation models, MCS-Net and DAC-Net, based on a symmetric six-level U-shaped architecture that enhances feature representation with low computational cost. The key innovation is the integration of attention mechanisms and multi-scale contextual cues into the UNet framework for more discriminative feature learning and improved boundary delineation. Experimental results show improved DSC and mIoU in skin lesion segmentation, demonstrating the effectiveness of the proposed designs.

Neural network-based medical image classification assists clinicians in identifying disease types and serves as a core technology for diagnostic support. Due to the high similarity between visual features and the limited labeled data in medical images, the performance of methods relying solely on visual feature extraction continues to be constrained. While preliminary research has investigated the incorporation of medical metadata to alleviate these problems, two critical issues persist: (1) ineffective modeling of the fine-grained semantic relationships between image and text, and (2) the computational overhead of efficient medical text encoding. To address these issues, we propose the Mamba-based Medical Metadata Net (TriMNet), a cross-modal fusion model grounded in State Space Models (SSM). We design a metadata encoder based on SSM, leveraging Mamba’s linear computational complexity and implicit state space modeling capabilities to enhance inference speed. Additionally, TriMNet incorporates a Cross-modal Gated Fusion Module (CGFM) to adaptively capture semantic associations between image regions and textual metadata. Furthermore, we introduce a modality-aware loss weighting strategy to improve the performance on rare disease types and enhance generalization across heterogeneous datasets. Experiments conducted on MedIMeta and PAD-UFES-20 demonstrate that TriMNet outperforms both single-modal and baseline models in classification accuracy. Notably, on PAD-UFES-20, our approach achieves a 1.4% improvement over the state-of-the-art.