Chinese Journal of Biomedical Engineering

Home

Editorial Board

Download

Quick Search

Office Online

Current Issue

Download Articles

Download

Links

2026 Vol. 45, No. 3 Published: 2026-06-20

Reviews Regular Papers

Regular Papers

	257	Research on Analysis Methods of Core Muscle Functional Networks During Hand Movements
		Du Yihao, Sun Mengyu, Li Jingjin, Wang Xiaoran, Wang Feng, Lin Shiguo, Wu Xiaoguang
		DOI: 10.3969/j.issn.0258-8021.2026.03.001
		To analyze the connectivity of the muscle functional network in hand movements and its relevance to tasks, revealing muscle synergy and its neuromodulatory mechanisms, hand-movement experimental paradigms (G&L experiment and G experiment) were designed. A-mode ultrasound signals and pressure signals were acquired using an ELONXI ultrasonic signal device and a self-developed hand pressure acquisition system. Twelve healthy subjects were recruited as research participants. Muscle functional network parameters (weighted node degree, weighted clustering coefficient, global efficiency) were calculated, and a non-negative tensor factorization model based on adaptive gradient descent (AGDSNTF) was employed to analyze the connectivity of muscle functional networks during hand movements. Moreover, through hand-pressure analysis and screening, the topological structure and dynamic characteristics of the functional network of key action points during hand movements were obtained. By analyzing the changes in muscle morphology and the differences in pressure distribution at key action points, the relevance between muscle synergy and tasks in hand movements was explored. The experimental results showed that there were significant differences in the connection characteristics of the muscle functional network under different experimental paradigms. Moreover, the muscle connections were closer and the synergistic effect was stronger under the G&L experimental paradigm. The functional network analysis of hand pressure and key action points revealed that the thumb[action points (1,1) and (1,3)] and the middle finger[action points (3,1) and (3,3)] played key roles in hand grasping and lifting movements [with pressure values reaching(7.0±0.8) N and (6.0±0.9) N respectively], while the ring finger [action point (4,2)] provided additional stability during hand grasping movements. Under the G&L experimental paradigm, the muscles needed to undergo greater morphological adjustments (FDS: 18.2 %±2.3 %, FCU: 16.5 %±2.1 %) to meet the anti-gravity requirements and achieved fine control through enhanced multi-muscle synergy. Under the G experimental paradigm, the pressure distribution at the key action points was more uniform, and the synergistic mode was relatively simple.
		2026 Vol. 45 (3): 257-266 [Abstract] ( 78 ) HTML (1 KB) PDF (6952 KB) ( 24 )

	267	A Left Ventricular Motion Estimation Method for CMR Cine Sequences Based on Shape-AttentionCascading
		Wang Yu, Sun Zheng, Zhang Nan ,
		DOI: 10.3969/j.issn.0258-8021.2026.03.002
		Cardiac magnetic resonance cine (CMR cine) is a key non-invasive technique for assessing cardiac motion. However, its images often suffer from motion artifacts and improper imaging parameters, leading to blurring that compromises the boundary localization accuracy of existing left ventricular motion estimation methods, thereby affecting the reliability of cardiac function measurements. To address this issue, this study proposed a shape-attention cascade structure, consisting of a basic module and a sequence module, both of which was connected in cascade. The basic module enhanced boundary feature perception through a shape flow layer, enabling pixel-wise error correction of the initial prediction. The sequence module ensured motion estimation coherence using a bidirectional motion attention layer. In addition, this study introduced a novel data augmentation method targeting boundary noise and designs a weighted Dice loss function that integrates pixel confidence and frame confidence. Experimental results demonstratd that the proposed method significantly improved the estimation accuracy for the left ventricular motion. On the ACDC dataset, the Hausdorff distance of the proposed method is 3.83 mm, outperforming the baseline model at 7.51 mm; on the private dataset, the proposed method achieves 4.39 mm, outperforming the baseline model at 8.68 mm; on EchoNet-Dynamic, the proposed method achieves 4.86 mm, outperforming the baseline method at 5.87 mm. The above differences were all statistically significant as assessed by the Kruskal-Wallis test(P<0.05). This study provides an effective method for accurate cardiac motion estimation, with positive implications for advancing the clinical translation of automated cardiac function assessment.
		2026 Vol. 45 (3): 267-275 [Abstract] ( 50 ) HTML (1 KB) PDF (6414 KB) ( 12 )

	276	Using Improved Transformer Model Considering Multi-domain Features for CT Image Recognitionof Connective Tissue-Associated Interstitial Lung Disease
		Gao Jing, Li Lei, Zhu Lingyan, Qian Tianhe, Wang Yongfu
		DOI: 10.3969/j.issn.0258-8021.2026.03.003
		Connectivetissue disease-related interstitial lung disease (CTD-ILD) is a type of chronic respiratory disease with an increasing number of patients. Early detection and diagnosis of this disease can effectively improve patients' treatment outcomes and survival rates. Pulmonary imaging is a primary method for the early diagnosis of CTD-ILD. To extract valuable information quickly from many medical images and determine the severity of the disease, an improved Transformer model considering multi-domain features was proposed in this work. This model was based on multi-domain feature collaborative learning, integrating the ResNet and Vision Transformer (ViT). It used the self-attention mechanism in the complex lesion area and depthwise separable convolution in the simple non-lesion area for feature extraction. The experimental dataset comprised 2,752 CTD-ILD lung CT images labeled by professional physicians. The model's validity was systematically evaluated through comparative experiments, ablation studies, and robustness assessments under noisy conditions. On the self-built dataset, the model achieved an accuracy of 96.71 %, precision of 96.74 %, recall of 96.71 %, and F1 score of 96.69 %, representing a 5.83 % accuracy improvement over ResNet50 and outperforming 9 classic models. The average single training epoch took 17 minutes, 5 minutes shorter than the unimproved model, with only 28.7 seconds required to train a single image. Through multi-domain feature fusion and optimization of the ViT architecture, the proposed method has demonstrated excellent performance in medical image recognition for CTD-ILD, achieving a dual improvement in the image recognition performance and training efficiency. This method is expected to assist clinicians in diagnosis and improve diagnostic efficiency.
		2026 Vol. 45 (3): 276-287 [Abstract] ( 71 ) HTML (1 KB) PDF (4140 KB) ( 21 )

	288	Deep Learning-Based Image-Level Detection of APL-Related Azurophilic Granule-PositiveLeukemia Cells
		Lin Zhiyuan, Liu Xin, Qiu Jinming, Mai Xiuqu, Zhang Fuhua, Xu Jiayu
		DOI: 10.3969/j.issn.0258-8021.2026.03.004
		The aim of this work is to develop a deep learning method for automatically identifying acute promyelocytic leukemia (APL)-related azurophilic granule-positive microscopic images, to improve the efficiency of morphological prescreening and support the early auxiliary diagnosis of APL in emergency clinical settings. A total of 450 leukocyte microscopic images collected from Nanhai People′s Hospital were used to construct an image-level binary classification dataset for the presence or absence of typical azurophilic granules, including 230 negative and 220 positive images. The dataset was divided into training and test sets at a ratio of 8∶2 using stratified random sampling. An end-to-end network, termed EADF-Net, was proposed based on EfficientNet-B4, which incorporated a feature pyramid attention module to capture multi-scale contextual information, a multi-scale feature fusion module to enhance fine-grained granule and boundary representations, and a context enhancement module to aggregate local and global contextual features. During training, binary cross-entropy loss, a Sobel gradient-based boundary enhancement regularization term, and supervised contrastive loss were jointly optimized. The proposed model was compared with ResNet-50, DenseNet-121, EfficientNet variants, and Swin Transformer under the same training settings. On the test set, EADF-Net achieved an AUC of 92.2%, accuracy of 90.0%, sensitivity of 84.1%, specificity of 95.7%, and F1-score of 89.1%. Compared with ResNet-50, EfficientNet-B4, EfficientNetV2-S, and Swin Transformer, EADF-Net showed better overall performance while maintaining a smaller model size. Ablation experiments further demonstrated that the combination of FPA, MSFF, and CEM effectively improved the model′s ability to discriminate fine-grained morphological features of APL. In conclusion, EADF-Net enabled stable discrimination of APL-related azurophilic granule-positive microscopic images under limited-sample conditions, while achieving improved performance with controlled model complexity, suggesting its potential as an auxiliary tool for early APL screening.
		2026 Vol. 45 (3): 288-298 [Abstract] ( 56 ) HTML (1 KB) PDF (5699 KB) ( 15 )

	299	Classification of Motor Imagery EEG Signals Based on MC-TANet
		Zhang Xuejun, Song Zhongchen
		DOI: 10.3969/j.issn.0258-8021.2026.03.005
		To address challenges in motor imagery EEG (MI-EEG) classification caused by low signal-to-noise ratio, non-stationarity, and significant individual differences, a classification model based on a multi-branch convolution and temporal attention network (MC-TANet) was proposed in this work. The model employed a multi-branch convolution module to extract temporal features in parallel using convolutional kernels of different scales. After that, the model integrated spatial and temporal information through depthwise spatial convolution and spatio-temporal convolution. After sliding window processing, a multi-head self-attention layer focuses on key features, and a temporal convolutional network mines deep spatio-temporal dependencies. These features were then fused via average fusion and fed into the classifier, where a fully connected layer and SoftMax perform the final probability prediction for the MI classification task. Experimental results on the BCI-2a and BCI-2b datasets showed that in subject-dependent experiments, the model achieved average accuracies of 87.9% and 90.8%, and Kappa coefficients of 0.838 and 0.816, respectively. In cross-subject experiments, accuracies of 69.6% and 80.3% and Kappa coefficients of 0.602 and 0.606 were achieved. It was confirmed that the model effectively captured mu and beta rhythms features, thereby enhancing inter-class separability.
		2026 Vol. 45 (3): 299-311 [Abstract] ( 61 ) HTML (1 KB) PDF (14569 KB) ( 13 )

	312	Research on Personalized FES Regulation Based on sEMG GDSNTF
		Du Yihao, Wang Xiaoran, Sun Mengyu, Li Jingjin, Wu Xiaoguang
		DOI: 10.3969/j.issn.0258-8021.2026.03.006
		Functional electrical stimulation (FES) is widely used in fields of motor recovery and neurological rehabilitation. To improve rehabilitation efficiency and achieve personalized rehabilitation, and to address the issues of real-time performance and stability in FES control, this study proposed a model based on gradient descent optimized L_1/2-norm sparsenon-negative tensor factorization (GDSNTF). This model simultaneously extracted muscle synergy information in three dimensions: temporal, spatial, and frequency domains, enabling a comprehensive and in-depth exploration of muscle synergy characteristics during human movement. The GDSNTF model was applied to the regulation of FES stimulation parameters. By introducing a weight optimization algorithm for primary motor patterns, the real-time performance and stability of FES control were improved, and weighted muscle synergies were used to enhance the personalization of FES stimulation. Ten healthy adult males were recruited as subjects, and a hand movement mirror control experiment was designed. Four movements were selected: wrist flexion, wrist extension, fist clenching, and palm opening. Surface electromyography (sEMG) signals were collected from the unaffected hand during movement, and muscle synergy features (temporal, spatial, and frequency domains) were extracted using GDSNTF to regulate FES stimulation parameters (current intensity, amplitude, and frequency) on the affected side. Comparative analyses were performed with non-negative tensor factorization (NTF) and sparse non-negative tensor factorization (SNTF). One-way analysis of variance (ANOVA) was used to verify the effectiveness of the personalized FES control method based on GDSNTF. The results showed that after 20 iterations, the relative errors of NTF and SNTF were 0.30±0.03 and 0.29±0.02, respectively, whereas that of GDSNTF was 0.20±0.02, with highly significant differences among groups (F=68.52, P<0.01), indicating higher accuracy of the GDSNTF algorithm. Meanwhile, the runtime of GDSNTF was (2.10±0.15) s, significantly shorter than (2.67±0.21) s for NTF and (2.48±0.18) s for SNTF, with highly significant inter-group differences (F=42.36, P<0.01). In conclusion, the proposed method achieved superior real-time performance while ensuring accuracy.
		2026 Vol. 45 (3): 312-320 [Abstract] ( 54 ) HTML (1 KB) PDF (6229 KB) ( 9 )

	321	Research on Cardiovascular Disease Prediction Model Based on Transformer
		Zhang Peilong, Zhao Feng, Bi Rigege, Si Yidan, Zhang Haoli, Tu Wa, Li Hua
		DOI: 10.3969/j.issn.0258-8021.2026.03.007
		Cardiovascular disease has become a major public health problem worldwide,and its prediction and early intervention are of great significance in reducing its morbidity and mortality. This study aimed to propose a Transformer-based cardiovascular disease prediction model to overcome the limitations of traditional models in modeling complex dependencies among multiple clinical features. The model used 13 physiological parameters (such as age, blood sugar, cholesterol, chest pain type, etc.) for standardization and encoding processing, and applied the self-attention mechanism of the Transformer to model the global dependencies among multidimensional clinical features, thereby improving the extraction of complex feature interactions.Theexperimentaldata were divided into a training set and a test set, and the modelwas trained and tested on the data. Experimental results showedthat the accuracy of the proposed Transformer modelwas 98.68% on the training set and 94.68% on the test set, whichwas 8% higher than the accuracy of traditional convolutional neural networks and deep neural networks. The loss rate of the Transformer model on the training set gradually tended to 0, which is significantly better than the loss rate of 0.13 of the CNN model and the loss rate of the DNN model of 0.22; the loss rate of the Transformer model on the test set tended to 0.1, whichwas also significantly better than that of the CNN model of 0.3 and that of the DNN model of 0.4.In conclusion, the cardiovascular disease prediction model based on Transformer made rapid diagnosis through easy-to-obtain clinical data with high accuracy,providing strong technical support for early intervention and precise treatment and displaying application prospects and practicability.
		2026 Vol. 45 (3): 321-331 [Abstract] ( 60 ) HTML (1 KB) PDF (5658 KB) ( 16 )

	332	An Automated Pose Accuracy Detection System for Laparoscopic Surgical Robots: Designand Verification
		Chen Xiu, Jin Lukai, Chen Huiming, Hong Wei, Tian Ruixue, Zhang Mingwei, Zhang Tianyi, ZhangPeiming
		DOI: 10.3969/j.issn.0258-8021.2026.03.008
		The pose accuracy of laparoscopic surgical robots are generally assessed through manual methods characterized by static, discrete, and poorly integrated processes with frequent human intervention, thus rendering them incapable of dynamic master-slave mapping tracking and inadequate for standardized testing demands. To address these shortcomings, this study designed an automated detection system based on an open architecture for evaluating the pose performance indicators of laparoscopic surgical robots. By establishing a central control core, the system deeply integrated the high-precision visual measurement technology with motion control of a six-axis collaborative robotic arm, achieving closed-loop coordination between the precise motion control and synchronous multi-target visual measurement, thereby ensuring synchronized detection of the master and slave ends. Under the management of the testing software, the system enabled fully automated testing of the core performance metrics. Test results indicated that the system achieved a master-slave operation position accuracy and repeatability of no more than ±1.0 mm, an attitude accuracy of no more than 4.0°, and an attitude repeatability of no more than ±1.0°, meeting the precision evaluation standards. Moreover, the detection efficiency was 50% higher than that of manual methods. The methodology employed in the proposed automated detection system was scientifically sound, significantly enhanced detection efficiency, and provided the reliable assurance for the safety and effectiveness of laparoscopic surgical robots.
		2026 Vol. 45 (3): 332-343 [Abstract] ( 49 ) HTML (1 KB) PDF (6827 KB) ( 9 )

	344	Study on Pull-Out Strength and Failure Modes of Three-Periodic Minimal Surface Bone Implants
		Xiang Zhipeng, Zeng Kai, Li Song, Xing Baoying, Li Xiaohu, Zhang Zihao
		DOI: 10.3969/j.issn.0258-8021.2026.03.009
		Due to its controllable physical properties, TPMS bone implants are widely used in bone scaffolding engineering, and the pull-out strength and failure form of surgical screws for component fastening in TPMS bone implants are important issues of concern in theoretical research and clinical application. In this paper, the extraction behavior of screws in Primitive and Gyroid TPMS structures was systematically studied by combining finite element simulation and experimental verification. Firstly, the finite element model of primitive and gyroid structures with different porosity (60% and 70%) and unit cell size (5.0 mm, 7.5 mm) were established, the screw extraction process was simulated, its mechanical response and failure mode were predicted, and the nylon PA12 specimen was prepared by additive manufacturing technology, and the extraction test was carried out to verify the simulation results. The results showed that the simulation was highly consistent with the experimental results, which verified the reliability of the model. Among them, the Gyroid structure showed better pull-out performance than the Primitive structure due to its continuous stress distribution characteristics, and the pull-out strength reached 1329.2 N under the optimal parameter combination (porosity of 60% and unit cell size of 7.5 mm). The study further revealed the relationship between the pull-out strength and the failure mode: when the overall block fracture occurred in the structure, the energy absorption value (> 150 mJ) and pull-out strength were high. Local fragmentation fractures corresponded to the lower energy absorption (< 150 mJ) and pull-out strength. The failure mode was regulated by the size of the unit cell, and the stability of the screw could be improved by adjusting this parameter to guide the structure to tend to break as a whole. This study clarified the influence of TPMS structural parameters on screw stability, and provided a theoretical basis and modeling method reference for the anti-extraction design and performance optimization of porous structures in bone implants.
		2026 Vol. 45 (3): 344-351 [Abstract] ( 56 ) HTML (1 KB) PDF (3895 KB) ( 12 )

	352	Fabrication of Puerarin@FeS₂/Ti₃C₂ Nanoplatform for Postoperative Bone Tumor Therapy
		Qing You , Li Suiyan, Weng Jie
		DOI: 10.3969/j.issn.0258-8021.2026.03.010
		Addressing the clinical challenge that current osteosarcoma treatment strategies struggle to simultaneously achieve tumor eradication and bone repair, this study designed a multifunctional nanomaterial to synchronously inhibiting osteosarcoma cells and protecting the proliferative activity of osteoblasts, thereby modulating homeostasis to facilitate bone defect repair. FeS₂ nanoparticles with excellent peroxidase-like activity were loaded onto the surface of Ti₃C₂ MXenes, which possess near-infrared (NIR) photothermal response properties, constructing an FeS₂/Ti₃C₂ composite carrier. The substrate was surface-functionalized using amino-polyethylene glycol-hydroxyl (NH₂-PEG-OH). The active herbal component Puerarin was immobilized onto the composite carrier via dithiodiglycolic acid bridging technology, ultimately yielding the Puerarin@FeS₂/Ti₃C₂ nanoplatform. Material characterization techniques including X-ray diffraction (XRD), scanning electron microscopy (SEM), transmission electron microscopy (TEM), Raman spectroscopy, and Fourier transform infrared spectroscopy (FT-IR) confirmed the successful preparation and composite structural features of the material. Cytotoxicity assays demonstrated that the material exhibited no significant toxicity to mouse fibroblast cells (L929) or macrophage cells (RAW264.7), with cell viability remaining over 80% after long-term co-culture. Under 1 064 nm NIR laser irradiation, the material exerts a photothermal-chemodynamic synergistic response. Compared with the single-stimulus groups, the composite material group exhibits enhanced inhibitory effects on osteosarcoma cell proliferation, with a UMR-106 cell viability of 54.3% after 2 days. At the same time, the introduction of Puerarin exerted a protective effect on BMSCs, showing no significant difference in the proliferation rate compared to the blank control group after 7 days. In conclusion, this work provided a novel strategy for developing integrated nanoplatforms for combined tumor therapy and bone defect repair.
		2026 Vol. 45 (3): 352-361 [Abstract] ( 55 ) HTML (1 KB) PDF (11107 KB) ( 7 )

Reviews

	362	Research Advances in Deep Learning-Based Multimodal Epilepsy Detection and Seizure Prediction
		Yu Jian, Deng Qishichao, Wu Jiacheng, Li Chuantao, Lin Dongmei, Yan Zuojian, Chen Fuming
		DOI: 10.3969/j.issn.0258-8021.2026.03.011
		Epilepsy is a chronic neurological disorder caused by abnormal synchronous neuronal discharge in the brain, affecting more than 70 million people worldwide. Accurate detection and seizure prediction of epilepsy are of critical clinical significance for reducing the risk of sudden unexpected death in epilepsy and improving patients' quality of life. Current mainstream methods mostly rely on single electroencephalogram (EEG) signals, which can hardly comprehensively capture the multi-system pathological changes accompanying epileptic seizures. This article reviewed the research progress in deep learning-based multimodal epilepsy detection and seizure prediction. First, commonly used physiological signals were summarized, including electroencephalogram (EEG), electrocardiogram (ECG), accelerometer and electrodermal activity, as well as public multimodal datasets such as EPILEPSIAE and TUSZ. Then, the principles, advantages, limitations and applicable scenarios of three core fusion strategies (early fusion, late fusion and hybrid fusion) were compared. The technical characteristics and performance of convolutional neural networks (CNN), long short-term memory networks (LSTM) and hybrid deep learning models were elaborated, and the mainstream technical route that takes EEG as the core and complements with other multimodal physiological signals were clarified. Finally, core challenges were pointed out, such as insufficient cross-patient generalization ability, difficulties in cross-modal alignment, non-standard data annotation and poor model interpretability. Future research trends were prospected from four aspects: developing robust cross-modal feature fusion algorithms, promoting standardized data construction, enhancing clinical interpretability and realizing lightweight deployment.
		2026 Vol. 45 (3): 362-371 [Abstract] ( 62 ) HTML (1 KB) PDF (3331 KB) ( 28 )

	372	Research Progress of Self-Powered Tumor Electrostimulation Therapy
		Liang Qilin, Liu Hui, Wu Di, Wang Ruixue, Zhao Chaochao
		DOI: 10.3969/j.issn.0258-8021.2026.03.012
		Malignant tumors have a high fatality rate and pose a serious threat to human health. Conventional surgical treatment, chemotherapy, and radiotherapy have certain drawbacks. As an emerging treatment method, tumor electrostimulation (ES) therapy has been clinically verified for its excellent safety and therapeutic effect, but energy supply insufficiency and equipment technology issues have hindered its further clinical application. With the rapid development of materials science, nanoscience, and micro-nano processing technology, self-powered technology has emerged, providing a new solution for tumor ES therapy. This review first introduced the working principle of self-powered technology, then discussed the research progress of self-powered tumor ES therapy, including self-powered electroporation (EP) therapy, self-powered electrochemical therapy (ECT), self-powered triboelectric immunotherapy (TIT), and self-powered dynamic therapy (DT). Finally,the latest work of self-powered technology in tumor detection and prevention of tumor metastasis was presented, along with core challenges in tumor therapy related to the mechanisms of electrical stimulation, treatment methods, devices and clinical application. The review also discussed the future development directions,to provide new ideas and strategies for tumor treatment.
		2026 Vol. 45 (3): 372-384 [Abstract] ( 58 ) HTML (1 KB) PDF (1992 KB) ( 9 )

Copyright © Editorial Board of Chinese Journal of Biomedical Engineering
Supported by:Beijing Magtech