生物医学工程学杂志最新文献

Cardiovascular diseases and psychological disorders represent two major threats to human physical and mental health. Research on electrocardiogram (ECG) signals offers valuable opportunities to address these issues. However, existing methods are constrained by limitations in understanding ECG features and transferring knowledge across tasks. To address these challenges, this study developed a multi-resolution feature encoding network based on residual networks, which effectively extracted local morphological features and global rhythm features of ECG signals, thereby enhancing feature representation. Furthermore, a model compression-based continual learning method was proposed, enabling the structured transfer of knowledge from simpler tasks to more complex ones, resulting in improved performance in downstream tasks. The multi-resolution learning model demonstrated superior or comparable performance to state-of-the-art algorithms across five datasets, including tasks such as ECG QRS complex detection, arrhythmia classification, and emotion classification. The continual learning method achieved significant improvements over conventional training approaches in cross-domain, cross-task, and incremental data scenarios. These results highlight the potential of the proposed method for effective cross-task knowledge transfer in ECG analysis and offer a new perspective for multi-task learning using ECG signals.

In audiovisual emotion recognition, representational learning is a research direction receiving considerable attention, and the key lies in constructing effective affective representations with both consistency and variability. However, there are still many challenges to accurately realize affective representations. For this reason, in this paper we proposed a cross-modal audiovisual recognition model based on a multi-head cross-attention mechanism. The model achieved fused feature and modality alignment through a multi-head cross-attention architecture, and adopted a segmented training strategy to cope with the modality missing problem. In addition, a unimodal auxiliary loss task was designed and shared parameters were used in order to preserve the independent information of each modality. Ultimately, the model achieved macro and micro F1 scores of 84.5% and 88.2%, respectively, on the crowdsourced annotated multimodal emotion dataset of actor performances (CREMA-D). The model in this paper can effectively capture intra- and inter-modal feature representations of audio and video modalities, and successfully solves the unity problem of the unimodal and multimodal emotion recognition frameworks, which provides a brand-new solution to the audiovisual emotion recognition.

Alzheimer's disease (AD) classification models usually segment the entire brain image into voxel blocks and assign them labels consistent with the entire image, but not every voxel block is closely related to the disease. To this end, an AD auxiliary diagnosis framework based on weakly supervised multi-instance learning (MIL) and multi-scale feature fusion is proposed, and the framework is designed from three aspects: within the voxel block, between voxel blocks, and high-confidence voxel blocks. First, a three-dimensional convolutional neural network was used to extract deep features within the voxel block; then the spatial correlation information between voxel blocks was captured through position encoding and attention mechanism; finally, high-confidence voxel blocks were selected and combined with multi-scale information fusion strategy to integrate key features for classification decision. The performance of the model was evaluated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Open Access Series of Imaging Studies (OASIS) datasets. Experimental results showed that the proposed framework improved ACC and AUC by 3% and 4% on average compared with other mainstream frameworks in the two tasks of AD classification and mild cognitive impairment conversion classification, and could find the key voxel blocks that trigger the disease, providing an effective basis for AD auxiliary diagnosis.

Medical cross-modal retrieval aims to achieve semantic similarity search between different modalities of medical cases, such as quickly locating relevant ultrasound images through ultrasound reports, or using ultrasound images to retrieve matching reports. However, existing medical cross-modal hash retrieval methods face significant challenges, including semantic and visual differences between modalities and the scalability issues of hash algorithms in handling large-scale data. To address these challenges, this paper proposes a Medical image Semantic Alignment Cross-modal Hashing based on Transformer (MSACH). The algorithm employed a segmented training strategy, combining modality feature extraction and hash function learning, effectively extracting low-dimensional features containing important semantic information. A Transformer encoder was used for cross-modal semantic learning. By introducing manifold similarity constraints, balance constraints, and a linear classification network constraint, the algorithm enhanced the discriminability of the hash codes. Experimental results demonstrated that the MSACH algorithm improved the mean average precision (MAP) by 11.8% and 12.8% on two datasets compared to traditional methods. The algorithm exhibits outstanding performance in enhancing retrieval accuracy and handling large-scale medical data, showing promising potential for practical applications.

This paper proposes a motor imagery recognition algorithm based on feature fusion and transfer adaptive boosting (TrAdaboost) to address the issue of low accuracy in motor imagery (MI) recognition across subjects, thereby increasing the reliability of MI-based brain-computer interfaces (BCI) for cross-individual use. Using the autoregressive model, power spectral density and discrete wavelet transform, time-frequency domain features of MI can be obtained, while the filter bank common spatial pattern is used to extract spatial domain features, and multi-scale dispersion entropy is employed to extract nonlinear features. The IV-2a dataset from the 4 ^th International BCI Competition was used for the binary classification task, with the pattern recognition model constructed by combining the improved TrAdaboost integrated learning algorithm with support vector machine (SVM), k nearest neighbor (KNN), and mind evolutionary algorithm-based back propagation (MEA-BP) neural network. The results show that the SVM-based TrAdaboost integrated learning algorithm has the best performance when 30% of the target domain instance data is migrated, with an average classification accuracy of 86.17%, a Kappa value of 0.723 3, and an AUC value of 0.849 8. These results suggest that the algorithm can be used to recognize MI signals across individuals, providing a new way to improve the generalization capability of BCI recognition models.

In order to realize the quantitative assessment of muscle strength in hand function rehabilitation and then formulate scientific and effective rehabilitation training strategies, this paper constructs a multi-scale convolutional neural network (MSCNN) - convolutional block attention module (CBAM) - bidirectional long short-term memory network (BiLSTM) muscle strength prediction model to fully explore the spatial and temporal features of the data and simultaneously suppress useless features, and finally achieve the improvement of the accuracy of the muscle strength prediction model. To verify the effectiveness of the model proposed in this paper, the model in this paper is compared with traditional models such as support vector machine (SVM), random forest (RF), convolutional neural network (CNN), CNN - squeeze excitation network (SENet), MSCNN-CBAM and MSCNN-BiLSTM, and the effect of muscle strength prediction by each model is investigated when the hand force application changes from 40% of the maximum voluntary contraction force (MVC) to 60% of the MVC. The research results show that as the hand force application increases, the effect of the muscle strength prediction model becomes worse. Then the ablation experiment is used to analyze the influence degree of each module on the muscle strength prediction result, and it is found that the CBAM module plays a key role in the model. Therefore, by using the model in this article, the accuracy of muscle strength prediction can be effectively improved, and the characteristics and laws of hand muscle activities can be deeply understood, providing assistance for further exploring the mechanism of hand functions .

In the clinical stage, suspected hemolytic plasma may cause hemolysis illness, manifesting as symptoms such as heart failure, severe anemia, etc. Applying a deep learning method to plasma images significantly improves recognition accuracy, so that this paper proposes a plasma quality detection model based on improved "You Only Look Once" 5th version (YOLOv5). Then the model presented in this paper and the evaluation system ‌were introduced‌ into the plasma datasets, and ‌the average accuracy of the final classification reached 98.7%‌. The results of this paper's experiment were obtained through the combination of several key algorithm modules including‌ omni-dimensional dynamic convolution, pooling with separable kernel attention, residual bi-fusion feature pyramid network, ‌and‌ re-parameterization convolution. The method of this paper‌ obtains the feature information of spatial mapping efficiently, and enhances the average recognition accuracy of plasma quality detection. This paper presents a high-efficiency detection method for plasma images, aiming to provide a practical approach to prevent hemolysis illnesses caused by external factors.

Synthetic CT (sCT) generated from CBCT has proven effective in artifact reduction and CT number correction, facilitating precise radiation dose calculation. However, the quality of different regions in sCT images is severely imbalanced, with soft tissue region exhibiting notably inferior quality compared to others. To address this imbalance, we proposed a Multi-Task Attention Network (MuTA-Net) based on VGG-16, specifically focusing the enhancement of image quality in soft tissue region of sCT. First, we introduced a multi-task learning strategy that divides the sCT generation task into three sub-tasks: global image generation, soft tissue region generation and bone region segmentation. This approach ensured the quality of overall sCT image while enhancing the network's focus on feature extraction and generation for soft tissues region. The result of bone region segmentation task guided the fusion of sub-tasks results. Then, we designed an attention module to further optimize feature extraction capabilities of the network. Finally, by employing a results fusion module, the results of three sub-tasks were integrated, generating a high-quality sCT image. Experimental results on head and neck CBCT demonstrated that the sCT images generated by the proposed MuTA-Net exhibited a 12.52% reduction in mean absolute error in soft tissue region, compared to the best performance among the three comparative methods, including ResNet, U-Net, and U-Net++. It can be seen that MuTA-Net is suitable for high-quality sCT image generation and has potential application value in the field of CBCT guided adaptive radiation therapy.

Riboflavin-ultraviolet A (UVA) collagen cross-linking has not only achieved good clinical efficacy in the treatment of corneal diseases such as dilatation keratopathy, bullae keratopathy, infectious keratopathy, and in the combined treatment of corneal refractive surgeries, but also its efficacy and safety in scleral collagen cross-linking have been initially confirmed. To better promote the application of cross-linking in the clinical treatment of corneal and scleral diseases, exploring controllability and predictability of the surgical efficacy are both important for evaluating the surgical efficacy and personalized precision treatment. In this paper, the progress on the cross-linking depth of riboflavin-UVA collagen cross-linking, and its relationship with the cross-linking effect will be reviewed. It will provide the reference for further application of this procedure in ophthalmology clinics.

Existing emotion recognition research is typically limited to static laboratory settings and has not fully handle the changes in emotional states in dynamic scenarios. To address this problem, this paper proposes a method for dynamic continuous emotion recognition based on electroencephalography (EEG) and eye movement signals. Firstly, an experimental paradigm was designed to cover six dynamic emotion transition scenarios including happy to calm, calm to happy, sad to calm, calm to sad, nervous to calm, and calm to nervous. EEG and eye movement data were collected simultaneously from 20 subjects to fill the gap in current multimodal dynamic continuous emotion datasets. In the valence-arousal two-dimensional space, emotion ratings for stimulus videos were performed every five seconds on a scale of 1 to 9, and dynamic continuous emotion labels were normalized. Subsequently, frequency band features were extracted from the preprocessed EEG and eye movement data. A cascade feature fusion approach was used to effectively combine EEG and eye movement features, generating an information-rich multimodal feature vector. This feature vector was input into four regression models including support vector regression with radial basis function kernel, decision tree, random forest, and K-nearest neighbors, to develop the dynamic continuous emotion recognition model. The results showed that the proposed method achieved the lowest mean square error for valence and arousal across the six dynamic continuous emotions. This approach can accurately recognize various emotion transitions in dynamic situations, offering higher accuracy and robustness compared to using either EEG or eye movement signals alone, making it well-suited for practical applications.