Biomedical Signal Processing and Control最新文献_第6页

Hybrid feature optimization and radial basis function networks for cardiovascular disease prediction 心血管疾病预测的混合特征优化与径向基函数网络

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109606

S. Savitha , A. Rajiv Kannan , K. Logeswaran

The study addresses the critical challenge of accurately predicting cardiovascular disease (CVD), a leading cause of mortality worldwide, where early diagnosis is crucial for effective intervention. Traditional models often struggle with high-dimensional data, imbalanced classes, and nonlinear feature interactions, limiting prediction reliability. Motivated by these gaps, this research proposes a hybrid methodology integrating Harris Hawks Search (HHS) for feature optimization with Radial Basis Function Networks (RBFN) to enhance CVD risk assessment. The HHS algorithm efficiently selects key predictive features such as chest pain type and number of vessels, reducing dimensionality while preserving vital information. Trained on optimized features, the RBFN classifier achieved superior performance with 92.1% accuracy, high sensitivity, and specificity, surpassing conventional models like Logistic Regression (81.2%) and Random Forest (86.7%). Ablation studies confirm each component’s contribution, with significant gains validated statistically (p < 0.05). The hybrid model also offers computational efficiency with training times around 31.7 s. Future work aims to validate this approach on diverse, larger datasets and integrate it into real-time clinical decision support systems, advancing personalized, interpretable, and efficient cardiovascular healthcare tools.

该研究解决了准确预测心血管疾病（CVD）的关键挑战，心血管疾病是全球死亡的主要原因，其中早期诊断对于有效干预至关重要。传统模型经常与高维数据、不平衡类和非线性特征交互作斗争，限制了预测的可靠性。基于这些不足，本研究提出了一种结合Harris Hawks搜索（HHS）和径向基函数网络（RBFN）进行特征优化的混合方法，以增强心血管疾病的风险评估。HHS算法有效地选择胸痛类型和血管数量等关键预测特征，在保留重要信息的同时降低了维数。经过优化的特征训练后，RBFN分类器的准确率达到92.1%，灵敏度和特异性都很高，超过了Logistic回归（81.2%）和随机森林（86.7%）等传统模型。消融研究证实了每个成分的贡献，统计学上证实了显著的获益（p < 0.05）。混合模型还提供了计算效率，训练时间约为31.7秒。未来的工作旨在在不同的、更大的数据集上验证这种方法，并将其整合到实时临床决策支持系统中，推进个性化、可解释和高效的心血管医疗工具。

{"title":"Hybrid feature optimization and radial basis function networks for cardiovascular disease prediction","authors":"S. Savitha , A. Rajiv Kannan , K. Logeswaran","doi":"10.1016/j.bspc.2026.109606","DOIUrl":"10.1016/j.bspc.2026.109606","url":null,"abstract":"<div><div>The study addresses the critical challenge of accurately predicting cardiovascular disease (CVD), a leading cause of mortality worldwide, where early diagnosis is crucial for effective intervention. Traditional models often struggle with high-dimensional data, imbalanced classes, and nonlinear feature interactions, limiting prediction reliability. Motivated by these gaps, this research proposes a hybrid methodology integrating Harris Hawks Search (HHS) for feature optimization with Radial Basis Function Networks (RBFN) to enhance CVD risk assessment. The HHS algorithm efficiently selects key predictive features such as chest pain type and number of vessels, reducing dimensionality while preserving vital information. Trained on optimized features, the RBFN classifier achieved superior performance with 92.1% accuracy, high sensitivity, and specificity, surpassing conventional models like Logistic Regression (81.2%) and Random Forest (86.7%). Ablation studies confirm each component’s contribution, with significant gains validated statistically (p < 0.05). The hybrid model also offers computational efficiency with training times around 31.7 s. Future work aims to validate this approach on diverse, larger datasets and integrate it into real-time clinical decision support systems, advancing personalized, interpretable, and efficient cardiovascular healthcare tools.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109606"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038921","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Flexible multi-modal classification network for endometrial carcinoma diagnosis 子宫内膜癌诊断的灵活多模态分类网络

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109574

Lingling Fang, Wenhui Zhang, Yongcheng Yu, Qian Wu

Accurate prognostic evaluation of endometrial cancer relies on precise pathological grading and lymphovascular space invasion (LVSI) assessment. To overcome the inherent limitation of single-modal MRI in characterizing heterogeneous tumors, this paper proposes a Flexible Multi-modal Classification Network (FMCNet). FMCNet effectively integrates complementary MRI sequences (T1WI, T2WI, ADC, DWI) through two novel modules: an Explicit-Implicit Feature analysis module, which hierarchically disentangles shallow anatomical and deep pathological features via adaptive recalibration, and a Cross-modal Enhancement Fusion framework, which constructs unified representations through attention-guided interactions to synergize complementary information while suppressing redundancy. Evaluated on a clinical dataset of 297 patients (2,889 images), FMCNet achieves state-of-the-art performance with an overall accuracy of 93.4% (94.8% for T1WI + T2WI + ADC), significantly outperforming conventional models (VGG16: 89.2%; DenseNet: 87.6%). Ablation studies confirm the critical roles of both modules. The framework’s ability to mitigate inter-modal interference underscores its potential for improving diagnostic assessment.

准确的子宫内膜癌预后评估依赖于精确的病理分级和淋巴血管间隙浸润（LVSI）评估。为了克服单模态MRI在表征异质性肿瘤方面的固有局限性，本文提出了一种灵活的多模态分类网络（FMCNet）。FMCNet通过两个新模块有效整合互补的MRI序列（T1WI, T2WI， ADC， DWI）：一个显式-隐式特征分析模块，通过自适应重新校准分层分离浅层解剖和深层病理特征；一个跨模态增强融合框架，通过注意引导交互构建统一表征，协同互补信息，同时抑制冗余。在297名患者（2889张图像）的临床数据集上进行评估，FMCNet达到了最先进的性能，总体准确率为93.4% (T1WI + T2WI + ADC为94.8%)，显著优于传统模型（VGG16: 89.2%; DenseNet: 87.6%）。消融研究证实了这两个模块的关键作用。该框架减轻多式联运干扰的能力强调了其改进诊断评估的潜力。

{"title":"Flexible multi-modal classification network for endometrial carcinoma diagnosis","authors":"Lingling Fang, Wenhui Zhang, Yongcheng Yu, Qian Wu","doi":"10.1016/j.bspc.2026.109574","DOIUrl":"10.1016/j.bspc.2026.109574","url":null,"abstract":"<div><div>Accurate prognostic evaluation of endometrial cancer relies on precise pathological grading and lymphovascular space invasion (LVSI) assessment. To overcome the inherent limitation of single-modal MRI in characterizing heterogeneous tumors, this paper proposes a Flexible Multi-modal Classification Network (FMCNet). FMCNet effectively integrates complementary MRI sequences (T1WI, T2WI, ADC, DWI) through two novel modules: an Explicit-Implicit Feature analysis module, which hierarchically disentangles shallow anatomical and deep pathological features via adaptive recalibration, and a Cross-modal Enhancement Fusion framework, which constructs unified representations through attention-guided interactions to synergize complementary information while suppressing redundancy. Evaluated on a clinical dataset of 297 patients (2,889 images), FMCNet achieves state-of-the-art performance with an overall accuracy of 93.4% (94.8% for T1WI + T2WI + ADC), significantly outperforming conventional models (VGG16: 89.2%; DenseNet: 87.6%). Ablation studies confirm the critical roles of both modules. The framework’s ability to mitigate inter-modal interference underscores its potential for improving diagnostic assessment.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109574"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038920","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Harmonizing heartbeats: RR-interval correction with Gaussian Process for reliable HRV 调和心跳：用高斯过程进行可靠HRV的rr -间隔校正

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109513

Hamed Mojtahed , Ramesh R. Rao , Christopher Paolini , Mahasweta Sarkar

Background and Objective:

Heart Rate Variability (HRV) measures derived from RR-intervals serve as crucial biomarkers for cardiovascular system responses to disease, physical activity, and stress. While healthcare professionals and general users utilize these measurements to gauge an individual’s well-being and assess their stress and fitness levels, their accuracy can be compromised by low-cost wearable devices, motion artifacts, poor electrode contact, and other factors that affect the RR-interval time series data derived from electrocardiograms. This study aims to develop and evaluate a Gaussian Process framework for correcting compromised RR-intervals.

Methods:

Monte Carlo simulations were used to distribute errors across the RR-interval series, which were sectioned by physical activity. Uniformly distributed additive Gaussian noise was synthetically introduced at various error rates (10%, 20%, and 30%) and durations (1 s, 3 s, 5 s, 7 s, and variable). Multiple Gaussian Process kernels, including basic and hybrid combinations, were applied to reconstruct the erroneous series. The results are compared with the widely used Cubic Spline Interpolation (CSI). Correlation analysis was employed to study the effect of error duration and presence levels on reconstruction ability.

Results:

The Gaussian process with a Rational Quadratic kernel demonstrated substantial improvements in deviation across all 27 HRV metrics by 87.66% and 90.26% compared to CSI for fixed-duration and variable-duration conditions. When compared with raw noisy signals, the reduction reached 94% for both conditions. These figures were further improved when combining Rational Quadratic and Periodic kernels, with deviation reductions of 88.22% (fixed-duration) and 90.77% (variable-duration) relative to CSI.

Conclusions:

The Gaussian Process framework, particularly with the RQ kernel, provides a reliable method for correcting compromised RR-interval measurements and recovering HRV variables with the least deviation for most parameters. This approach could enhance the reliability of HRV measurements from consumer-grade devices, potentially improving both clinical assessments and personal health monitoring.

背景和目的：心率变异性（HRV）测量来源于rr间隔，是心血管系统对疾病、身体活动和压力反应的重要生物标志物。虽然医疗保健专业人员和一般用户使用这些测量来衡量个人的健康状况并评估他们的压力和健康水平，但它们的准确性可能会受到低成本可穿戴设备、运动伪影、不良电极接触以及其他影响心电图衍生的rr间隔时间序列数据的因素的影响。本研究旨在开发和评估一个高斯过程框架，用于纠正受损的rr区间。方法：采用蒙特卡罗模拟方法在rr区间序列中分布误差，该区间序列由体育活动分割。在不同错误率（10%、20%和30%）和持续时间（1 s、3 s、5 s、7 s和变量）下，综合引入均匀分布的加性高斯噪声。采用多高斯过程核，包括基本和混合组合，重建错误序列。结果与广泛使用的三次样条插值（CSI）进行了比较。采用相关分析研究误差持续时间和存在水平对重建能力的影响。结果：与固定持续时间和可变持续时间条件下的CSI相比，具有理性二次核的高斯过程在所有27个HRV指标上的偏差均有显著改善，分别为87.66%和90.26%。与原始噪声信号相比，两种情况下的降幅均达到94%。当将有理二次核和周期核结合使用时，这些数据得到了进一步的改善，相对于CSI，在固定持续时间下偏差降低了88.22%，在可变持续时间下偏差降低了90.77%。结论：高斯过程框架，特别是RQ核，提供了一种可靠的方法来纠正受损的rr区间测量和恢复HRV变量，对大多数参数的偏差最小。这种方法可以提高消费者级设备HRV测量的可靠性，潜在地改善临床评估和个人健康监测。

{"title":"Harmonizing heartbeats: RR-interval correction with Gaussian Process for reliable HRV","authors":"Hamed Mojtahed , Ramesh R. Rao , Christopher Paolini , Mahasweta Sarkar","doi":"10.1016/j.bspc.2026.109513","DOIUrl":"10.1016/j.bspc.2026.109513","url":null,"abstract":"<div><h3>Background and Objective:</h3><div>Heart Rate Variability (HRV) measures derived from RR-intervals serve as crucial biomarkers for cardiovascular system responses to disease, physical activity, and stress. While healthcare professionals and general users utilize these measurements to gauge an individual’s well-being and assess their stress and fitness levels, their accuracy can be compromised by low-cost wearable devices, motion artifacts, poor electrode contact, and other factors that affect the RR-interval time series data derived from electrocardiograms. This study aims to develop and evaluate a Gaussian Process framework for correcting compromised RR-intervals.</div></div><div><h3>Methods:</h3><div>Monte Carlo simulations were used to distribute errors across the RR-interval series, which were sectioned by physical activity. Uniformly distributed additive Gaussian noise was synthetically introduced at various error rates (10%, 20%, and 30%) and durations (1 s, 3 s, 5 s, 7 s, and variable). Multiple Gaussian Process kernels, including basic and hybrid combinations, were applied to reconstruct the erroneous series. The results are compared with the widely used Cubic Spline Interpolation (CSI). Correlation analysis was employed to study the effect of error duration and presence levels on reconstruction ability.</div></div><div><h3>Results:</h3><div>The Gaussian process with a Rational Quadratic kernel demonstrated substantial improvements in deviation across all 27 HRV metrics by 87.66% and 90.26% compared to CSI for fixed-duration and variable-duration conditions. When compared with raw noisy signals, the reduction reached 94% for both conditions. These figures were further improved when combining Rational Quadratic and Periodic kernels, with deviation reductions of 88.22% (fixed-duration) and 90.77% (variable-duration) relative to CSI.</div></div><div><h3>Conclusions:</h3><div>The Gaussian Process framework, particularly with the RQ kernel, provides a reliable method for correcting compromised RR-interval measurements and recovering HRV variables with the least deviation for most parameters. This approach could enhance the reliability of HRV measurements from consumer-grade devices, potentially improving both clinical assessments and personal health monitoring.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109513"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038919","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Optimized eyeblink artifact removal in EEG signal using attention-based spiking neural networks 基于注意的脉冲神经网络优化脑电信号中眨眼伪影的去除

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109571

M. Sarathkumar , J. Rajalakshmi , Jeyapandi Marimuthu , A. Solairaj

Electroencephalogram (EEG) is widely used to monitor the cerebral activities. However, artifacts arising from non-cerebral sources such as eye movements, eye blinks and muscle activities often corrupt the recorded signals and significantly reducing reliability. To address this challenge, this paper proposes an Optimized Eyeblink Artifact Removal in EEG using Attention-Based Spiking Neural Networks (EAR-ASNN-OSP). The proposed framework integrates four core modules: Signed Cumulative Distribution Transform (SCDT) for robust feature extraction, Attention Spiking Neural Network (ASNN) for artifact detection, Efficient Binary Crayfish Optimization (EBCO) for adaptive weight parameter optimization, and a Dendritic Neural Network (DNN) for effective artifact removal. The SCDT captures discriminative features such as variance, average rectified value, and peak-to-peak amplitude, which are then supplied to the ASNN for classification of EEG signals into normal, corrupt, and eyeblink categories. EBCO further enhances detection performance by optimizing the ASNN weights, while the DNN ensures accurate removal of identified artifacts. The system was implemented in Python, and its performance was rigorously evaluated using metrics such as accuracy, precision, recall, F1-score, entropy, and computational time. Experimental results demonstrate that EAR-ASNN-OSP outperforms existing techniques including AEMD-RES-KCLN, QEF-ML-EAD, and DBP-EWT-EES. Specifically, the proposed method achieves up to 15.17 % higher accuracy, 20.11 % higher precision, and 19.19 % higher F1-score compared to state-of-the-art approaches. These findings highlight the effectiveness and robustness of EAR-ASNN-OSP, making it a promising solution for reliable EEG artifact detection and removal, thereby enhancing the quality and interpretability of EEG-based clinical and research applications.

脑电图（EEG）被广泛用于监测大脑活动。然而，眼球运动、眨眼和肌肉活动等非大脑来源产生的伪影经常会破坏记录的信号，并大大降低可靠性。为了解决这一挑战，本文提出了一种基于注意力的峰值神经网络（EAR-ASNN-OSP）的优化的EEG眨眼伪迹去除方法。该框架集成了四个核心模块：用于鲁棒特征提取的签名累积分布变换（SCDT）、用于伪迹检测的注意力峰值神经网络（ASNN）、用于自适应权重参数优化的高效二进制小龙虾优化（EBCO）和用于有效去除伪迹的树突状神经网络（DNN）。SCDT捕获判别特征，如方差、平均整流值和峰对峰幅度，然后将其提供给ASNN，用于将EEG信号分类为正常、损坏和眨眼类别。EBCO通过优化ASNN权重进一步提高检测性能，而DNN则确保准确去除已识别的伪像。该系统是用Python实现的，其性能通过准确性、精密度、召回率、f1分数、熵和计算时间等指标进行了严格评估。实验结果表明，EAR-ASNN-OSP优于aemd - re - kcln、QEF-ML-EAD和DBP-EWT-EES等现有技术。具体而言，与现有方法相比，该方法的准确率提高了15.17%，精密度提高了20.11%，f1分数提高了19.19%。这些发现突出了EAR-ASNN-OSP的有效性和鲁棒性，使其成为可靠的脑电信号伪迹检测和去除的有希望的解决方案，从而提高了基于脑电图的临床和研究应用的质量和可解释性。

{"title":"Optimized eyeblink artifact removal in EEG signal using attention-based spiking neural networks","authors":"M. Sarathkumar , J. Rajalakshmi , Jeyapandi Marimuthu , A. Solairaj","doi":"10.1016/j.bspc.2026.109571","DOIUrl":"10.1016/j.bspc.2026.109571","url":null,"abstract":"<div><div>Electroencephalogram (EEG) is widely used to monitor the cerebral activities. However, artifacts arising from non-cerebral sources such as eye movements, eye blinks and muscle activities often corrupt the recorded signals and significantly reducing reliability. To address this challenge, this paper proposes an Optimized Eyeblink Artifact Removal in EEG using Attention-Based Spiking Neural Networks (EAR-ASNN-OSP). The proposed framework integrates four core modules: Signed Cumulative Distribution Transform (SCDT) for robust feature extraction, Attention Spiking Neural Network (ASNN) for artifact detection, Efficient Binary Crayfish Optimization (EBCO) for adaptive weight parameter optimization, and a Dendritic Neural Network (DNN) for effective artifact removal. The SCDT captures discriminative features such as variance, average rectified value, and peak-to-peak amplitude, which are then supplied to the ASNN for classification of EEG signals into normal, corrupt, and eyeblink categories. EBCO further enhances detection performance by optimizing the ASNN weights, while the DNN ensures accurate removal of identified artifacts. The system was implemented in Python, and its performance was rigorously evaluated using metrics such as accuracy, precision, recall, F1-score, entropy, and computational time. Experimental results demonstrate that EAR-ASNN-OSP outperforms existing techniques including AEMD-RES-KCLN, QEF-ML-EAD, and DBP-EWT-EES. Specifically, the proposed method achieves up to 15.17 % higher accuracy, 20.11 % higher precision, and 19.19 % higher F1-score compared to state-of-the-art approaches. These findings highlight the effectiveness and robustness of EAR-ASNN-OSP, making it a promising solution for reliable EEG artifact detection and removal, thereby enhancing the quality and interpretability of EEG-based clinical and research applications.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109571"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038922","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Advanced detection of myositis muscle images based on regional enhanced deep learning models integrated with SVM and image processing techniques 基于支持向量机和图像处理技术的区域增强深度学习模型的肌炎图像高级检测

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109490

Wessam M. Salama

A hybrid deep learning–machine learning framework for the automatic classification of myositis muscle illnesses from ultrasound images is presented in this paper. The suggested approach combines a Support Vector Machine (SVM) classifier and a regionally improved Fast R-CNN architecture with a VGG16 backbone to take advantage of robust decision boundaries and deep feature extraction. To increase the representation of muscle texture, a thorough preprocessing pipeline comprising edge-based noise suppression, contrast enhancement, and median filtering is put into place. A five-fold cross-validation approach is used to guarantee dependability, and performance metrics are presented with 95% Confidence Intervals (CI) and standard deviation. Strong discriminative ability is demonstrated by Receiver Operating Characteristic (ROC) analysis, which produced consistent results across folds and a high average AUC of 98.32%. The suggested framework outperforms current deep architectures with the greatest classification accuracy of 98.46%, according to comparative tests against ResNet50, DenseNet121, and EfficientNet-B0. These results demonstrate the stability, repeatability, and clinical application of the model for fine-grained distinction between the classes of Inclusion Body Myositis (IBM), Polymyositis (PM), Dermatomyositis (DM), and Normal. Therefore, the suggested Fast R-CNN–VGG16–SVM methodology offers a computationally effective and statistically verified method for automated myositis detection utilizing ultrasound imaging.

本文提出了一种用于超声图像中肌炎肌肉疾病自动分类的混合深度学习-机器学习框架。该方法将支持向量机（SVM）分类器和区域改进的Fast R-CNN架构与VGG16主干相结合，以利用鲁棒决策边界和深度特征提取。为了增强肌肉纹理的表征，我们采用了一套完整的预处理流程，包括基于边缘的噪声抑制、对比度增强和中值滤波。采用五重交叉验证方法来保证可靠性，性能指标以95%置信区间（CI）和标准差呈现。受试者工作特征（Receiver Operating Characteristic， ROC）分析显示了较强的判别能力，各褶皱的结果一致，平均AUC高达98.32%。根据与ResNet50、DenseNet121和EfficientNet-B0的比较测试，建议的框架以98.46%的最高分类准确率优于当前的深度架构。这些结果证明了该模型在包络体肌炎（IBM）、多发性肌炎（PM）、皮肌炎（DM）和正常类别之间精细区分的稳定性、可重复性和临床应用。因此，建议的Fast R-CNN-VGG16-SVM方法为利用超声成像自动检测肌炎提供了一种计算有效且经过统计验证的方法。

{"title":"Advanced detection of myositis muscle images based on regional enhanced deep learning models integrated with SVM and image processing techniques","authors":"Wessam M. Salama","doi":"10.1016/j.bspc.2026.109490","DOIUrl":"10.1016/j.bspc.2026.109490","url":null,"abstract":"<div><div>A hybrid deep learning–machine learning framework for the automatic classification of myositis muscle illnesses from ultrasound images is presented in this paper. The suggested approach combines a Support Vector Machine (SVM) classifier and a regionally improved Fast R-CNN architecture with a VGG16 backbone to take advantage of robust decision boundaries and deep feature extraction. To increase the representation of muscle texture, a thorough preprocessing pipeline comprising edge-based noise suppression, contrast enhancement, and median filtering is put into place. A five-fold cross-validation approach is used to guarantee dependability, and performance metrics are presented with 95% Confidence Intervals (CI) and standard deviation. Strong discriminative ability is demonstrated by Receiver Operating Characteristic (ROC) analysis, which produced consistent results across folds and a high average AUC of 98.32%. The suggested framework outperforms current deep architectures with the greatest classification accuracy of 98.46%, according to comparative tests against ResNet50, DenseNet121, and EfficientNet-B0. These results demonstrate the stability, repeatability, and clinical application of the model for fine-grained distinction between the classes of Inclusion Body Myositis (IBM), Polymyositis (PM), Dermatomyositis (DM), and Normal. Therefore, the suggested Fast R-CNN–VGG16–SVM methodology offers a computationally effective and statistically verified method for automated myositis detection utilizing ultrasound imaging.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109490"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039276","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multimodal Brownian bridge diffusion model for controllable synthetic medical image generation 可控合成医学图像生成的多模态布朗桥扩散模型

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109584

Yu Li , Jiaqing Liu , Rahul Kumar Jain , Yen-Wei Chen

Deep learning requires large and diverse datasets to effectively learn downstream tasks. However, medical imaging often suffers from limited data availability, privacy and low diversity. Generative data augmentation (GDA) offers a promising solution by synthesizing labeled samples to expand training datasets and support downstream tasks. While recent advances in generative models have shown strong potential, many existing GDA approaches rely only on lesion annotation masks, often failing to preserve fine anatomical details or capture subtle morphological variations. To address these limitations, we propose a multimodal (mask- and text-guided) diffusion framework that enables fine-grained semantic control and maintains anatomical consistency. We construct detailed text datasets for medical images using a multimodal large language model. Further, the semantic features, such as lesion color, contour and background context, are integrated into the diffusion model employing an attention-based module. Our method maintains anatomical consistency while enabling diverse semantic variations through fine-grained control, achieved via proposed attention conditional batch normalization. Experiments on the Kvasir-SEG, ISIC 2016 and 3D-IRCADb-01 datasets demonstrate that our method significantly improves segmentation performance across various architectures. Code, datasets and pretrained models are available.¹

深度学习需要大量不同的数据集来有效地学习下游任务。然而，医学成像经常受到数据可用性有限、隐私和多样性低的困扰。生成数据增强（GDA）通过合成标记样本来扩展训练数据集并支持下游任务，提供了一种很有前途的解决方案。虽然生成模型的最新进展显示出强大的潜力，但许多现有的GDA方法仅依赖于病变注释掩模，往往无法保留精细的解剖细节或捕捉微妙的形态变化。为了解决这些限制，我们提出了一个多模态（掩码和文本引导）扩散框架，该框架可以实现细粒度的语义控制并保持解剖一致性。我们使用多模态大语言模型构建医学图像的详细文本数据集。此外，利用基于注意力的模块将病灶颜色、轮廓和背景上下文等语义特征集成到扩散模型中。我们的方法保持解剖一致性，同时通过细粒度控制实现不同的语义变化，通过提出的注意力条件批处理规范化实现。在Kvasir-SEG、ISIC 2016和3D-IRCADb-01数据集上的实验表明，我们的方法显著提高了不同架构的分割性能。代码、数据集和预训练模型都是可用的

{"title":"Multimodal Brownian bridge diffusion model for controllable synthetic medical image generation","authors":"Yu Li , Jiaqing Liu , Rahul Kumar Jain , Yen-Wei Chen","doi":"10.1016/j.bspc.2026.109584","DOIUrl":"10.1016/j.bspc.2026.109584","url":null,"abstract":"<div><div>Deep learning requires large and diverse datasets to effectively learn downstream tasks. However, medical imaging often suffers from limited data availability, privacy and low diversity. Generative data augmentation (GDA) offers a promising solution by synthesizing labeled samples to expand training datasets and support downstream tasks. While recent advances in generative models have shown strong potential, many existing GDA approaches rely only on lesion annotation masks, often failing to preserve fine anatomical details or capture subtle morphological variations. To address these limitations, we propose a multimodal (mask- and text-guided) diffusion framework that enables fine-grained semantic control and maintains anatomical consistency. We construct detailed text datasets for medical images using a multimodal large language model. Further, the semantic features, such as lesion color, contour and background context, are integrated into the diffusion model employing an attention-based module. Our method maintains anatomical consistency while enabling diverse semantic variations through fine-grained control, achieved via proposed attention conditional batch normalization. Experiments on the Kvasir-SEG, ISIC 2016 and 3D-IRCADb-01 datasets demonstrate that our method significantly improves segmentation performance across various architectures. Code, datasets and pretrained models are available.<span><span><sup>1</sup></span></span></div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109584"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038917","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Cardiovascular disease detection using google wide slice residual network approach by electrocardiogram images 基于谷歌宽切片残差网络的心电图图像心血管疾病检测

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109530

Dugumari Siva Raja Kumar , Balajee Maram , Pravin Ramdas Kshirsagar , Telagarapu Prabhakar

Cardiovascular Disease (CVD) refers to a collection of heart and blood vessel disorders caused by the build-up of plaque in the arteries, which restricts the blood flow throughout the body. Existing detection methods failed to identify the disease at an early stage and may cause discomfort during certain diagnostic procedures. Hence, the DL-based Google Wide Slice Residual Network (G-WISeR-Net) model is introduced for effective CVD detection. The proposed G-WISeR-Net model employs the four most pivotal subprocesses: denoising, feature extraction, feature fusion, and detection. Initially, ECG images are sourced from the database, and binary image conversion-based denoising is performed. Then, medical features, Stationary Wavelet Transform (SWT), and statistical features, like mean, variance, relative energy, relative amplitude, entropy, kurtosis, and information gain, are extracted. Subsequently, the extracted features are fused by a Deep Residual Network (DRN) with Topsoe similarity. Then, CVD detection is performed by G-WISeR-Net, which is an integration of Google Network (GoogleNet) and Wide Slice Residual Network (WISeR). The experimental results highlight that the proposed G-WISeR-Net attained an accuracy of 91.645%, sensitivity of 92.786%, and specificity of 91.479%.

心血管疾病（CVD）是指由动脉斑块积聚引起的心脏和血管疾病，这限制了血液在全身的流动。现有的检测方法未能在早期阶段识别疾病，并可能在某些诊断过程中引起不适。因此，引入了基于dl的谷歌宽片残余网络（g - wise - net）模型，用于有效的CVD检测。提出的g - wise - net模型采用了四个最关键的子过程：去噪、特征提取、特征融合和检测。首先，从数据库中获取心电图像，并进行基于二值图像转换的去噪。然后，提取医学特征、平稳小波变换（SWT）以及均值、方差、相对能量、相对幅值、熵、峰度和信息增益等统计特征。随后，利用Topsoe相似度的深度残差网络（DRN）对提取的特征进行融合。然后，CVD检测由g - wise - net进行，该网络是谷歌网络（GoogleNet）和Wide Slice Residual Network （WISeR）的集成。实验结果表明，G-WISeR-Net的准确率为91.645%，灵敏度为92.786%，特异性为91.479%。

{"title":"Cardiovascular disease detection using google wide slice residual network approach by electrocardiogram images","authors":"Dugumari Siva Raja Kumar , Balajee Maram , Pravin Ramdas Kshirsagar , Telagarapu Prabhakar","doi":"10.1016/j.bspc.2026.109530","DOIUrl":"10.1016/j.bspc.2026.109530","url":null,"abstract":"<div><div>Cardiovascular Disease (CVD) refers to a collection of heart and blood vessel disorders caused by the build-up of plaque in the arteries, which restricts the blood flow throughout the body. Existing detection methods failed to identify the disease at an early stage and may cause discomfort during certain diagnostic procedures. Hence, the DL-based Google Wide Slice Residual Network (G-WISeR-Net) model is introduced for effective CVD detection. The proposed G-WISeR-Net model employs the four most pivotal subprocesses: denoising, feature extraction, feature fusion, and detection. Initially, ECG images are sourced from the database, and binary image conversion-based denoising is performed. Then, medical features, Stationary Wavelet Transform (SWT), and statistical features, like mean, variance, relative energy, relative amplitude, entropy, kurtosis, and information gain, are extracted. Subsequently, the extracted features are fused by a Deep Residual Network (DRN) with Topsoe similarity. Then, CVD detection is performed by G-WISeR-Net, which is an integration of Google Network (GoogleNet) and Wide Slice Residual Network (WISeR). The experimental results highlight that the proposed G-WISeR-Net attained an accuracy of 91.645%, sensitivity of 92.786%, and specificity of 91.479%.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109530"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039386","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Exploring the challenge and value of deep learning in automated skin disease diagnosis 探讨深度学习在皮肤病自动诊断中的挑战和价值

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109589

Runhao Liu , Ziming Chen , Guangzhen Yao , Peng Zhang

Skin cancer is one of the most prevalent and deadly forms of cancer worldwide, highlighting the critical importance of early detection and diagnosis in improving patient outcomes. Deep learning (DL) has shown significant promise in enhancing the accuracy and efficiency of automated skin disease diagnosis, particularly in detecting and classifying skin lesions. However, several challenges remain for DL-based skin cancer diagnosis, including complex features, image noise, intra-class variation, inter-class similarity, and data imbalance. This review synthesizes recent research and discusses innovative approaches to address these challenges, such as data augmentation, hybrid models, and feature fusion. Furthermore, the review highlights the integration of DL models into clinical workflows, offering insights into the potential of deep learning to revolutionize skin disease diagnosis and improve clinical decision-making. This review uniquely integrates a PRISMA-based methodology with a challenge-oriented taxonomy, providing a systematic and transparent synthesis of recent deep learning advances for skin disease diagnosis. It further highlights emerging directions such as hybrid CNN-Transformer architectures and uncertainty-aware models, emphasizing its contribution to future dermatological AI research.

皮肤癌是世界上最普遍和最致命的癌症之一，这突出了早期发现和诊断对改善患者预后的至关重要性。深度学习（DL）在提高自动皮肤病诊断的准确性和效率方面显示出巨大的希望，特别是在检测和分类皮肤病变方面。然而，基于dl的皮肤癌诊断仍然存在一些挑战，包括复杂的特征、图像噪声、类内差异、类间相似性和数据不平衡。本文综合了最近的研究，并讨论了应对这些挑战的创新方法，如数据增强、混合模型和特征融合。此外，该综述强调了将深度学习模型集成到临床工作流程中，为深度学习在皮肤病诊断和改善临床决策方面的潜力提供了见解。本综述独特地将基于prisma的方法与面向挑战的分类法相结合，为皮肤疾病诊断提供了最近深度学习进展的系统和透明的综合。它进一步强调了新兴方向，如混合CNN-Transformer架构和不确定性感知模型，强调其对未来皮肤病学人工智能研究的贡献。

{"title":"Exploring the challenge and value of deep learning in automated skin disease diagnosis","authors":"Runhao Liu , Ziming Chen , Guangzhen Yao , Peng Zhang","doi":"10.1016/j.bspc.2026.109589","DOIUrl":"10.1016/j.bspc.2026.109589","url":null,"abstract":"<div><div>Skin cancer is one of the most prevalent and deadly forms of cancer worldwide, highlighting the critical importance of early detection and diagnosis in improving patient outcomes. Deep learning (DL) has shown significant promise in enhancing the accuracy and efficiency of automated skin disease diagnosis, particularly in detecting and classifying skin lesions. However, several challenges remain for DL-based skin cancer diagnosis, including complex features, image noise, intra-class variation, inter-class similarity, and data imbalance. This review synthesizes recent research and discusses innovative approaches to address these challenges, such as data augmentation, hybrid models, and feature fusion. Furthermore, the review highlights the integration of DL models into clinical workflows, offering insights into the potential of deep learning to revolutionize skin disease diagnosis and improve clinical decision-making. This review uniquely integrates a PRISMA-based methodology with a challenge-oriented taxonomy, providing a systematic and transparent synthesis of recent deep learning advances for skin disease diagnosis. It further highlights emerging directions such as hybrid CNN-Transformer architectures and uncertainty-aware models, emphasizing its contribution to future dermatological AI research.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109589"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039390","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A comprehensive review of robotics and deep learning applications during the COVID-19 pandemic 2019冠状病毒病疫情期间机器人和深度学习应用综述

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109459

S. Rooban , Iwin Thanakumar Joseph S , Mohamed Uvaze Ahamed Ayoobkhan , Muthukumara Rajaguru Kattiakara Muni Samy

The people are impacted by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) virus pandemic in worldwide over last two years, commonly known as coronavirus disease (COVID 19). In an effort to stop outbreak, governments of numerous nations have implemented measures like social distance and full or partial lockdowns. There is less interaction between people as a result of these COVID-19 preventive and control measures. It reviews the contribution of robotics, enhanced with deep learning techniques, to the challenges set during the COVID-19 pandemic. Attention is placed mainly on robotic applications for diagnosis, monitoring, disinfection, patient assistance, and rehabilitation; IoT and blockchain technologies are considered only when they directly support a robotic function. Existing research is analyzed in a structured manner to identify the performance, capabilities, and limitations of different robotic models and deep learning techniques. Comparisons among the different approaches are also made to provide a better understanding of the current trends and future directions. The findings point out that integration of robotics with deep learning enhances safety, efficiency, and healthcare support in pandemic conditions.

在过去的两年里，人们受到了严重急性呼吸综合征冠状病毒2 （SARS-CoV-2）病毒大流行的影响，通常被称为冠状病毒病（COVID - 19）。为了阻止疫情爆发，许多国家的政府都采取了社交距离、全面或部分封锁等措施。这些COVID-19预防和控制措施减少了人与人之间的互动。报告回顾了深度学习技术增强的机器人技术对2019冠状病毒病大流行期间所面临挑战的贡献。主要关注机器人在诊断、监测、消毒、病人协助和康复方面的应用；物联网和区块链技术只有在直接支持机器人功能时才被考虑。现有的研究以结构化的方式进行分析，以确定不同机器人模型和深度学习技术的性能、能力和局限性。还对不同的方法进行了比较，以便更好地了解当前的趋势和未来的方向。研究结果指出，将机器人技术与深度学习相结合，可以提高流行病条件下的安全性、效率和医疗保健支持。

{"title":"A comprehensive review of robotics and deep learning applications during the COVID-19 pandemic","authors":"S. Rooban , Iwin Thanakumar Joseph S , Mohamed Uvaze Ahamed Ayoobkhan , Muthukumara Rajaguru Kattiakara Muni Samy","doi":"10.1016/j.bspc.2026.109459","DOIUrl":"10.1016/j.bspc.2026.109459","url":null,"abstract":"<div><div>The people are impacted by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) virus pandemic in worldwide over last two years, commonly known as coronavirus disease (COVID 19). In an effort to stop outbreak, governments of numerous nations have implemented measures like social distance and full or partial lockdowns. There is less interaction between people as a result of these COVID-19 preventive and control measures. It reviews the contribution of robotics, enhanced with deep learning techniques, to the challenges set during the COVID-19 pandemic. Attention is placed mainly on robotic applications for diagnosis, monitoring, disinfection, patient assistance, and rehabilitation; IoT and blockchain technologies are considered only when they directly support a robotic function. Existing research is analyzed in a structured manner to identify the performance, capabilities, and limitations of different robotic models and deep learning techniques. Comparisons among the different approaches are also made to provide a better understanding of the current trends and future directions. The findings point out that integration of robotics with deep learning enhances safety, efficiency, and healthcare support in pandemic conditions.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109459"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039391","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A dual-branch multiscale model based on Bi-Mamba for EEG emotion recognition 基于Bi-Mamba的双分支多尺度脑电情绪识别模型

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-19 DOI: 10.1016/j.bspc.2026.109623

Lin Feng , Yongzhen Huo , Yuqiu Kong , Cheng Cheng

Emotion recognition using electroencephalography (EEG) has emerged as a key noninvasive physiological modality for human–computer interaction, intelligent healthcare, and mental-health monitoring. However, existing approaches often treat temporal and spatial features of EEG signals in isolation and lack bidirectional, deep interactive fusion mechanisms. Furthermore, multiscale analysis of temporal and spatial information remains limited in scope. To overcome these challenges, we introduce a Dual-Branch Multiscale Bi-Mamba framework (DBM-BiMamba). DBM-BiMamba is a unified, multiscale, bidirectional, spatiotemporal fusion framework. It comprises three components that are optimized together: a Multiscale Temporal Feature Learning (MTFL) branch, a Hierarchical Spatial Feature Learning (HSFL) branch and a Bidirectional Fusion Mamba (Bi-Mamba) module. The MTFL branch uses parallel depthwise separable convolutions, followed by a transformer encoder, to capture local micro-dynamics and long-range temporal dependencies. The HSFL branch performs hierarchical graph convolutions with node-wise attention to emphasize critical inter-electrode relationships and produce discriminative topological embeddings. The Bi-Mamba module then applies forward and backward state-space modeling to the temporal and spatial embeddings, fusing them at the sequence level and enabling efficient bidirectional spatiotemporal interactions rather than simply concatenating the features. Extensive ten-fold subject-independent cross-validation in the DEAP, SEED, and SEED-IV datasets demonstrates state-of-the-art accuracy of 95.55% for arousal and 95.02% for valence in DEAP, 94.77% in SEED, and 89.68% in SEED-IV. Ablation studies confirm the complementary gains of each component.

利用脑电图（EEG）进行情绪识别已成为人机交互、智能医疗和心理健康监测的一种关键的无创生理模式。然而，现有方法往往孤立地处理脑电信号的时空特征，缺乏双向、深度的交互融合机制。此外，时空信息的多尺度分析范围仍然有限。为了克服这些挑战，我们引入了双分支多尺度Bi-Mamba框架（DBM-BiMamba）。DBM-BiMamba是一个统一的、多尺度的、双向的、时空融合的框架。它包括三个共同优化的组件：一个多尺度时间特征学习（MTFL）分支，一个分层空间特征学习（HSFL）分支和一个双向融合曼巴（Bi-Mamba）模块。MTFL分支使用并行深度可分离卷积，然后是变压器编码器，以捕获局部微动力学和长期时间依赖性。HSFL分支使用节点关注执行分层图卷积，以强调关键的电极间关系并产生判别拓扑嵌入。然后，Bi-Mamba模块将前向和后向状态空间建模应用于时间和空间嵌入，在序列级别融合它们，实现有效的双向时空交互，而不是简单地将特征连接起来。对DEAP、SEED和SEED- iv数据集进行了广泛的十倍受试者独立交叉验证，结果显示，DEAP的唤醒准确率为95.55%，效价准确率为95.02%，SEED为94.77%，SEED- iv为89.68%。消融研究证实了每个部分的互补增益。

{"title":"A dual-branch multiscale model based on Bi-Mamba for EEG emotion recognition","authors":"Lin Feng , Yongzhen Huo , Yuqiu Kong , Cheng Cheng","doi":"10.1016/j.bspc.2026.109623","DOIUrl":"10.1016/j.bspc.2026.109623","url":null,"abstract":"<div><div>Emotion recognition using electroencephalography (EEG) has emerged as a key noninvasive physiological modality for human–computer interaction, intelligent healthcare, and mental-health monitoring. However, existing approaches often treat temporal and spatial features of EEG signals in isolation and lack bidirectional, deep interactive fusion mechanisms. Furthermore, multiscale analysis of temporal and spatial information remains limited in scope. To overcome these challenges, we introduce a Dual-Branch Multiscale Bi-Mamba framework (DBM-BiMamba). DBM-BiMamba is a unified, multiscale, bidirectional, spatiotemporal fusion framework. It comprises three components that are optimized together: a Multiscale Temporal Feature Learning (MTFL) branch, a Hierarchical Spatial Feature Learning (HSFL) branch and a Bidirectional Fusion Mamba (Bi-Mamba) module. The MTFL branch uses parallel depthwise separable convolutions, followed by a transformer encoder, to capture local micro-dynamics and long-range temporal dependencies. The HSFL branch performs hierarchical graph convolutions with node-wise attention to emphasize critical inter-electrode relationships and produce discriminative topological embeddings. The Bi-Mamba module then applies forward and backward state-space modeling to the temporal and spatial embeddings, fusing them at the sequence level and enabling efficient bidirectional spatiotemporal interactions rather than simply concatenating the features. Extensive ten-fold subject-independent cross-validation in the DEAP, SEED, and SEED-IV datasets demonstrates state-of-the-art accuracy of 95.55% for arousal and 95.02% for valence in DEAP, 94.77% in SEED, and 89.68% in SEED-IV. Ablation studies confirm the complementary gains of each component.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109623"},"PeriodicalIF":4.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039392","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0