Chinese Journal of Biomedical Engineering

Home

Editorial Board

Download

Quick Search

Office Online

Current Issue

Download Articles

Download

Links

2025 Vol. 44, No. 4 Published: 2025-08-20

Reviews Regular Papers

Regular Papers

	385	A Double-Stream Twin Contrast Network for EEG Emotion Recognition
		Ma Yuliang, Xie Yunzhen, Meng Ming, Gao Yunyuan, She Qingshan
		DOI: 10.3969/j.issn.0258-8021.2025.04.001
		In recent years, the research of emotion recognition based on EEG signals has gradually made remarkable progress. However, the labeling of labels requires a lot of manpower, and it is difficult to quickly obtain a large number of labeled data in practical applications. Therefore, how to utilize limited labels efficiently in the emotion recognition research become one bottleneck problem to overcome. In this work, a model architecture based on self-supervised double-flow twin network was proposed, which consisted of two interacting and learning branches of convolutional neural networks. First, the model was pre-trained. The amplified data of the input signal after two random signal transformations were input into the training branch and the target branch of the twin network. After extracting features from the convolutional module and the fully connected module in the branch, the model learned the general representation of the EEG signal in the process. Finally, the encoder part of the training branch was retained, and then the fully connected layer was used to fine-tune the model with labeled data, and the classification results are obtained. Data samples from public data sets SEED and SEED-IV were used to verify and evaluate the model classification effect. Under the fully labeled data, the classification accuracy of 93.92% and 89.71% were achieved, respectively. Under 50% label usage, the average accuracy of the three categories was 92.68%, which was only 1.24% less than that using all labels. The results showed that the model effectively extracted the general representation of EEG data, and achieved high recognition accuracy with relatively less labels.
		2025 Vol. 44 (4): 385-392 [Abstract] ( 81 ) HTML (1 KB) PDF (2295 KB) ( 37 )

	393	Graph Attention Mechanism-Based EEG Emotion Recognition Method
		Lu Weikun, Zhou Yingyue, Wu Qiao, Liao xiang, Liu Qi, Huang Runxia, Yang Bo
		DOI: 10.3969/j.issn.0258-8021.2025.04.002
		Emotion recognition plays a crucial role in human-computer interaction, and EEG signals have advantages in reflecting human emotional states. Modeling the complex interactions between brain regions based on the spatial topology of EEG can provide essential information for feature extraction in emotion recognition. However, using the single graph structure based on spatial topology for graph convolution suffers from the limitation of a single information aggregation method, making it difficult to describe the complex relationships between EEG channels. To address this issue, we proposed an EEG emotion recognition model based on a graph attention mechanism. The core idea was to use a multi-head graph attention mechanism to dynamically assign weights to node connections, thereby adaptively capturing the relationships between EEG channels and overcoming the shortcomings of existing methods in information aggregation and dynamic pattern capture. The proposed model was validated using 7 424 EEG samples from the DEAP dataset, achieving classification accuracies of 96.06%, 96.54%, and 96.84% on the three emotional dimensions of valence, arousal, and dominance, respectively. Compared to the ELGCNN model, which is also based on graph neural networks, the proposed model demonstrated improvements of 6.20% and 6.56% in the accuracy for the valence and arousal dimensions, proving its effectiveness. Additionally, the model′s performance was further validated using 4 630 EEG samples from the DREAMER dataset, yielding classification accuracies of 87.87%, 83.47%, and 79.96% for the valence, arousal, and dominance dimensions, respectively. The experimental results demonstrate that employing graph attention mechanisms can effectively extracted theemotion-related EEG features and improved the accuracy of EEG-based emotion recognition.
		2025 Vol. 44 (4): 393-405 [Abstract] ( 71 ) HTML (1 KB) PDF (3500 KB) ( 50 )

	406	Cross Domain Automatic Sleep Staging Based on Single Channel EEG
		Zhu Mengyuan, Quan Ruomeng, Qiang Ning, Hu Jing, Feng Feilong, Li Jin
		DOI: 10.3969/j.issn.0258-8021.2025.04.003
		Sleep staging is of great significance for diagnosis and treatment of sleep problems. Currently, existing deep sleep staging networks have limitations including low data efficiency and data distribution differences, resulting in a decrease in the model′s performance trained on the actual data from training sets. To address this problem, this paper proposed an adversarial domain adaptation network for cross-domain automatic sleep staging using single-channel EEG. This network used a feature extractor to extract EEG features, and at the same time used a non-shared attention mechanism to preserve key information of specific domains, combined a domain discriminator to align the source and target domains, and solved the class-level alignment problem through a dual-stage classifier based on iterative self-training. To verify the reliability of the proposed network, 39, 42, and 44 records were randomly selected from three public datasets of Sleep-EDF-20, SHHS1, and SHHS2 respectively. Experiments conducted on the six generated cross-domain scenarios showed that the proposed network achieved an average accuracy of 74.29% and an average MF1 value of 61.95%. When compared with the performance of other baseline models, the average accuracy of the proposed network was at least 2.01% higher than that of the existing baseline models, and the average MF1 score was at least 2.22% higher. This method provided a solution for addressing the domain shift problem in the sleep staging task.
		2025 Vol. 44 (4): 406-415 [Abstract] ( 62 ) HTML (1 KB) PDF (5081 KB) ( 26 )

	416	Cardiac Disease Classification Based on 3D Matrix Features Utilizing Multi-Channel Heart Sound Signals
		Fang Yu, Guo Zijian, Leng Hongxia, Liu Xing, Wang Weibo, Liu Dongbo, Wu Xiaochen
		DOI: 10.3969/j.issn.0258-8021.2025.04.004
		This study proposes a method for classifying cardiac diseases by extracting 3D matrix features from multi-channel heart sound signals, addressing the limitations of traditional methods that primarily utilize one-dimensional features from single-channel signals, which may overlook pathological correlations across different channels and cycles. First, a Butterworth filter was applied for noise reduction on the heart sound signals from each channel. The R-wave peaks were then located to segment the heart sounds, from which 15 effective time-frequency features, including Welch method power spectral energy, are extracted. Subsequently, these features were stacked into a 3D matrix with dimensions corresponding to the number of channels, cycles, and features, with the optimal cycle number determined to be 4. This 3D matrix was directly used as input for a CNN classifier. The method was tested on a dataset comprising 126 normal and 185 abnormal heart sounds, achieving an accuracy of 98.9%. Additionally, the method was validated on 126 clinical cases of congenital heart disease sounds and normal sounds, resulting in a classification accuracy of 93.9%. These experimental results indicated that the 3D matrix features could effectively capture pathological characteristics in the heart sound signals, improving the classification accuracy by 2.7% compared to single-channel features, providing valuable assistance for clinical cardiac treatment.
		2025 Vol. 44 (4): 416-423 [Abstract] ( 52 ) HTML (1 KB) PDF (3762 KB) ( 13 )

	424	Deep Graph Attention Networks with Memory for Prediction of Response to Neoadjuvant Chemotherapy in Breast Cancer
		Xu Weichao, Fan Ming, Fei Sixiang, Li Lihua
		DOI: 10.3969/j.issn.0258-8021.2025.04.005
		Neoadjuvant chemotherapy (NACT) is widely used in the treatment of breast cancer patients. Studies have shown that patients with a good chemotherapy response, particularly those achieving pathological complete response (pCR), have significantly improved survival rates. However, due to individual differences, some patients may experience poor responses or disease progression, therefore, early prediction of the response to NACT is crucial. Traditional convolutional neural networks (CNNs) have limitations in handling the spatial correlations and heterogeneity of tumor regions, which affects prediction accuracy. To address this, we proposed a graph network method combined with a deep graph attention network (DeepGAT) model and memory layers. We used the SLIC superpixel algorithm to segment the breast region and extract radiomic features from these superpixels. The superpixels were set as feature nodes, and edges were defined based on the weighted Euclidean distance and Pearson correlation coefficient of node features. By combining nodes and edges, we built a complete graph network. Our model included a DeepGAT module that dynamically weights node features using an attention mechanism to capture both local and global information. Additionally, we employed a memory pooling module based on clustering methods, which enhanced classification performance by capturing critical information and long-term dependencies. We used 214 samples, with 149 for training and 65 for testing. We also investigated the impact of graph sparsity coefficients on response to NACT performance, which showed that the best performance was achieved at a sparsity coefficient of 0.09. At this sparsity level, our model achieved AUCs of 0.822±0.027 and 0.811±0.041, significantly outperforming traditional GCN (AUC=0.726±0.045) and GAT (AUC=0.803±0.037) models. These results demonstrated the superior accuracy of our model in predicting response to NACT, providing a solid foundation for future research.
		2025 Vol. 44 (4): 424-434 [Abstract] ( 52 ) HTML (1 KB) PDF (2508 KB) ( 22 )

	435	Muscle Ultrasound Image Segmentation Based on Unsupervised Deep Learning Optical Flow Method
		Li Yang, Li Jin, Luan Kuan
		DOI: 10.3969/j.issn.0258-8021.2025.04.006
		Accurate image segmentation is the foundation of image guided treatment for spastic muscles after stroke. We proposed an unsupervised deep learning optical flow algorithm, aiming to segment the moving target muscle. Due to lack of labels of ultrasound image dataset of motor muscles not like the previous static ultrasound image datasets, an unsupervised optical flow estimation framework was constructed. The optical flow was learned from the unlabeled image sequence through augemented self-monitoring. At the same time, in order to avoid the influence of view synthesis targets on accuracy of augmented data, another forward propagation was added to the converted image to distort the basic learning framework, in which the conversion prediction from the original image is monitored, and finally, Gaussian filter, speed threshold filtering, border detection filtering and convex hull segmentation were employed to extract moving muscles from the optical flow field. The proposed method was evaluated on MPI Sintel, KITTI2012, KITTI2015, Flying Chairs and CityScapes (totally including 6 000 basic samples and 3 600 multi frame model samples) for its basic performance and cross dataset generalization performance, and muscle ultrsound image dataset containing 600 samples for its transfer performance was tested. The results demonstrated that the proposed method effectively extracted the dense optical flow field information without the use of optical flow labels in the dataset. The average endpoint error reached 5.80, the parameter quantity was reduced to 2.35 M, and good cross dataset generalization ability was obsersed. In the transfer test, compared with the manual segmentation results, the mean of center point differences was less than 1 mm, and the mean of intersection over unions was higher than 0.9. The proposedmethod can effectively segment the moving musclesof ultrasound images.
		2025 Vol. 44 (4): 435-446 [Abstract] ( 53 ) HTML (1 KB) PDF (9875 KB) ( 13 )

	447	Cascading Feature Fusion Polyp Segmentation Network Based on Efficient Additive Attention
		Li Meng, Zhang Sunjie
		DOI: 10.3969/j.issn.0258-8021.2025.04.007
		To address the problems of insufficient information interaction between local and global and weak correlation of features between neighboring layers at different depths in most polyp segmentation methods, this paper proposed a new network model (PVT-SMCD) based on the Pyramid Vision Transformer and self-attention mechanism cascaded decoder. Firstly, PVTv2 was used as the backbone network to extract image features, and the key information was obtained through the efficient additive attention module to capture the long-distance dependency relationships. Secondly, a multiple kernel convolution enhance blocks were introduced to locate high-level semantic features of polyps, and the obtained featureswere inputted to the cascade decoder to achieve the information interaction between the local and global layers. And lastly, a feature fusion module was used to gradually fuse the features between the two neighboring layers from top to bottom to reduce the information gap between the fused high-dimensional features and the low-dimensional features. The model in this paperwas compared with other eight medical image segmentation networks on five polyp segmentation datasets, mDice on Kvasir and CVC-ClinicDB datasets were 92.3% and 94.5%, mIoU were 87.1% and 89.9%, and the MAE were 0.021 and 0.006, respectively; on CVC-300, mDice and mIoU were 90% and 83.3%, respectively, with an MAE of 0.007; on CVC-ColonDB mDice were81.5%, mIoU was 73.5%, and MAE was 0.028; And on the ETIS dataset, mDice was 78.9%, mIoU was 71.3%, and MAE was0.019. Experimental results indicated that PVT-SMCD achieved superior performance over state-of-the-art methods across most of the evaluation metrics, demonstrating enhanced learning ability and generalization capacity, leading to more precise polyp segmentation outcomes.
		2025 Vol. 44 (4): 447-456 [Abstract] ( 47 ) HTML (1 KB) PDF (2509 KB) ( 14 )

	457	Long Term Animal Experimental Study on a Heterogeneous Tricuspid Annuloplasty Ring Madeof a Hyperelastic Nitinol
		Jia Liujun, Yue Guangxin, He Ting, Jiang Xin, Zhang Qi, Fan Junying, Song Jiangping
		DOI: 10.3969/j.issn.0258-8021.2025.04.008
		To evaluate whether a heterogeneous 3D tricuspid annuloplasty ring made of a hyperelastic nitinol can work effectively for a long time after implantation and to reveal its biocompatibility, 12 experimental samples (PerMed rings) and 9 control samples (Edward MC3 rings) were implanted into 21 small tailed Han sheep under circulation bypass. Pre-implantation, and at post implantation day (PID) 90 and PID 180, body weight was taken to assess animal growth, blood was collected for complete blood count, blood chemistries and free hemoglobin test to monitor infection and observe the interaction between the “rings” and blood, and the tricuspid valve function, cardiac structural changes, and thrombus formation was evaluated through transthoracic echocardiography (TTE). Then the animals were euthanized according to the plan, the integrity and local pathological changes, the local inflammatory response level and endothelialization of the prosthesis were evaluated by autopsy, histopathology, and SEM. Animals of both groups had good weight gain after surgery, with BW reaching (51.7±4.5) kg and (55.5±5.2) kg at post implantation day 180. Complete blood count showed there was no infection. Both types of prostheses did not cause hematolysis, abnormal changes in blood cells or hepatic/nephric dysfunction. TTE revealed that TVPG of animals from each group below 5 mmHg throughout the experiment, the tricuspid valves all worked normally, and no thrombus was detected. The pathology results showed all prostheses remained intact structures, closely adhered to the recipient tricuspid annulus, and did not break, crack, or deform, both types of rings had intact fibrous sacs on their surfaces. The local inflammatory response scores were (4.7±1.5) and (5.4±2.2) and the endothelialization rates were (74%±38%) and (82%±29%), respectively at post implantation day 180. The above results indicated that the heterogeneous tricuspid annuloplasty ring made of a hyperelastic Nitinol maintained tricuspid valve function safely and effectively for a long time, and with good blood compatibility and biocompatibility after being implanted into the small tailed Han sheep.
		2025 Vol. 44 (4): 457-464 [Abstract] ( 52 ) HTML (1 KB) PDF (6267 KB) ( 16 )

Reviews

	465	A Review of EEG Feature Extraction for Emotion Recognition Tasks
		Li Mengmeng, Xue Wenbo, Liu Yunyang, He Yudie, Yue Caitong, Li Zhihui, Shang Zhigang
		DOI: 10.3969/j.issn.0258-8021.2025.04.009
		Emotion plays a crucial role in human-computer interaction because it makes psychological and physiological responses to the environment. Accurate emotion recognition is vital for applications in the fields of medicine, education, psychology, and military. Compared to non-physiological signals, such as facial expressions and movements, physiological signals are difficult to disguise. As a type of physiological signal, electroencephalogram (EEG) offers advantages in terms of collectionconvenience and recognition accuracy and thus is often used in the field of emotion recognition. This review summarized recent progresses made in the feature extraction for EEG-based emotion recognition, introduced conventional time-frequency features, spatial domain features, brain network features, shallow nonlinear and manifold features, as well as deep learning-based feature extraction methods. Furthermore, this review provided an outlook on future directions.
		2025 Vol. 44 (4): 465-477 [Abstract] ( 67 ) HTML (1 KB) PDF (1167 KB) ( 94 )

	478	Research and Applications of Attention-Enhanced Generative Adversarial Network in MedicalImage Generation
		Fan Shanhui, Liang Shuxin, Wang Zhiwen, Wei Kaihua, Wang Qiang, Li Lihua
		DOI: 10.3969/j.issn.0258-8021.2025.04.010
		Generative adversarial network (GAN) have been demonstrated great research value and application potential in medical image generation, such as data enhancement and image quality improvement, owing to their excellent generation capabilities. However, traditional GAN models still face core challenges, including insufficient robustness and limited generalization capabilities. To address these issues, attention mechanisms, leveraging their strengths in modeling global feature correlations and focusing on key regions, offer a new technical pathway for enhancing GAN-based medical image generation. Thus, exploring how to efficiently combine attention mechanisms with GAN to boost generation quality has become a research hotspot in areas including denoising, reconstruction, data enhancement, and cross-modal generation. This paper systematically reviewed the advancements in attention-enhanced GAN techniques for medical image generation over the past five years (2019-2024). First, the principles of classical GAN and mainstream attention modules were introduced. Next, from a task-driven perspective, we critically analyzed the improved effects of attention mechanisms on GAN in different tasks. Finally, we delved into the current challenges and proposed potential future research directions. Through multidimensional analysis and discussion, this review hopes to provide valuable insights for advancing technologies in medical dataset expansion and image quality enhancement.
		2025 Vol. 44 (4): 478-493 [Abstract] ( 40 ) HTML (1 KB) PDF (1410 KB) ( 81 )

	494	Advances in the Application of 3D Printed Microneedle Biosensor for Continuous Glucose Monitoring
		Liu Jun, Fan Kai, Yao Danyu, Gan Kaifeng, SuChang, Wang Ling, Xu Mingen
		DOI: 10.3969/j.issn.0258-8021.2025.04.011
		Diabetes is a metabolic disease characterized by chronic hyperglycemia. Continuous glucose monitoring is a crucial component in the management of diabetes. Microneedle technology is an emerging method for blood glucose monitoring, which offers significant advantages such as painless, continuous, and real-time monitoring. Traditional microneedle manufacturing techniques, such as micromolding, have limitations including lower precision and lack of customizability. 3D printing technology, with its high precision, customization, and material versatility, has emerged as a new approach for microneedle fabrication. This review summarized the latest advancements in 3D printed microneedles, covering key technologies, printing principles, printable materials, and medical applications. We discussed the progress of microneedle-based glucose sensors in continuous glucose monitoring, including the underlying principles, microneedle preparation, electrochemical performance characterization, and wearable integration technologies. Finally, this review explored the challenges of 3D printed microneedles, along with future directions for the development of microneedle biosensors.
		2025 Vol. 44 (4): 494-501 [Abstract] ( 54 ) HTML (1 KB) PDF (1398 KB) ( 56 )

	502	Research Progress on Application of Molecularly Imprinted Electrochemical Sensors in BiomedicalField
		Shi Fan, Zhang Guojun, Li Yutao, Tang Lina
		DOI: 10.3969/j.issn.0258-8021.2025.04.012
		Molecular blotting technology mimics the antigen-antibody specific recognition mechanism of the organism's immune system, and designs and synthesises functional polymers with pre-assembled recognition sites to achieve highly selective recognition and capture of target molecules. In recent years, this technology has demonstrated its potential application in biomedical detection through cross-fertilisation with electrochemical sensing technology, and has attracted much attention for its excellent detection sensitivity, selectivity and stability. In this paper, we systematically review the construction elements of molecularly imprinted electrochemical sensors, focusing on the categories of functional monomers and their selection strategies, as well as the preparation of highly efficient imprinted membranes, and review the progress of the application of such sensors in biomedical detection fields such as biological and clinical marker detection and pathogen identification, and discuss the challenges of the development of sensors in the areas of sensitivity, precision, reproducibility, and suitability for complex biological samples. It also discusses the challenges of sensor development, such as sensitivity, precision, repeatability, and adaptation of complex biological samples, and looks forward to the future development of this field in the application of new smart materials, multimodal sensing integration and standardised detection platform construction, which will provide an important reference to promote the application of molecularly imprinted electrochemical sensors in the field of precision medicine.
		2025 Vol. 44 (4): 502-512 [Abstract] ( 48 ) HTML (1 KB) PDF (7760 KB) ( 47 )

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