Biomedical Signal Processing and Control最新文献_第7页

L2HiEMG synthesizer: A CycleGAN-driven computational tool for synthesizing enhanced muscle dynamics from lower-level intramuscular electromyography signals L2HiEMG合成器：cyclegan驱动的计算工具，用于从低水平肌内肌电图信号合成增强的肌肉动力学

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-20 DOI: 10.1016/j.bspc.2025.109364

Honghan Li , Jinyang Yu , Haijun Wu , Miguel Bordallo López , Yongkun Zhao

Intramuscular electromyography (iEMG) recorded at low contraction levels often exhibits overlapping signal characteristics, particularly within the range of mild to moderate muscle activations. This overlap reduces the discriminability of contraction states in the 25%–50% maximum voluntary contraction (MVC) range, making it challenging for iEMG-based prediction systems to achieve reliable classification. As a consequence, such systems frequently produce intermittent or discontinuous outputs, which compromise the naturalness and stability of prosthetic and exoskeleton control. To address this limitation, we propose the L2HiEMG synthesizer, a neural network model built upon the CycleGAN framework. The model is designed to transform low-intensity iEMG signals (25% MVC) into their higher-intensity counterparts (50% MVC), thereby simulating the nonlinear dynamics of motor unit recruitment and muscle fiber activation. Leveraging the dual Generator–Discriminator structure of CycleGAN and its cycle-consistency constraint, the proposed method effectively addresses issues arising from unpaired data, sample imbalance, and ambiguous label correspondence. Experimental evaluations demonstrated that the L2HiEMG synthesizer can generate high-fidelity signals at higher intensity. In the time domain, the synthesized signals exhibited mean and standard deviation differences of less than 0.01 and 0.04, respectively, with correlation coefficients consistently exceeding 0.99. Frequency-domain analyses further validated the accuracy of the generated signals, showing minimal deviations in zero-crossing rate (

<

0.08), signal energy (

<

0.01), and power spectral density (

<

0.03). Taken together, these results confirmed the model’s capability to produce physiologically realistic high-intensity iEMG signals, thereby offering a promising strategy to enhance the stability and responsiveness of iEMG-driven assistive control systems.

在低收缩水平记录的肌内肌电图（iEMG）经常显示重叠信号特征，特别是在轻度至中度肌肉激活范围内。这种重叠降低了25%-50%最大自主收缩（MVC）范围内收缩状态的可判别性，使得基于iemg的预测系统难以实现可靠的分类。因此，这样的系统经常产生间歇性或不连续的输出，这损害了假肢和外骨骼控制的自然性和稳定性。为了解决这一限制，我们提出了L2HiEMG合成器，这是一个建立在CycleGAN框架上的神经网络模型。该模型旨在将低强度iEMG信号（25% MVC）转换为高强度iEMG信号（50% MVC），从而模拟运动单元招募和肌纤维激活的非线性动力学。利用CycleGAN的双生成器-鉴别器结构及其循环一致性约束，该方法有效地解决了数据不配对、样本不平衡和标签对应不明确等问题。实验结果表明，该L2HiEMG合成器可以在较高的强度下产生高保真度的信号。在时域上，合成信号的均值和标准差差异分别小于0.01和0.04，相关系数均大于0.99。频域分析进一步验证了生成信号的准确性，表明过零率（<0.08）、信号能量（<0.01）和功率谱密度（<0.03）的偏差最小。综上所述，这些结果证实了该模型能够产生生理上真实的高强度iEMG信号，从而为增强iEMG驱动的辅助控制系统的稳定性和响应性提供了一种有希望的策略。

{"title":"L2HiEMG synthesizer: A CycleGAN-driven computational tool for synthesizing enhanced muscle dynamics from lower-level intramuscular electromyography signals","authors":"Honghan Li , Jinyang Yu , Haijun Wu , Miguel Bordallo López , Yongkun Zhao","doi":"10.1016/j.bspc.2025.109364","DOIUrl":"10.1016/j.bspc.2025.109364","url":null,"abstract":"<div><div>Intramuscular electromyography (iEMG) recorded at low contraction levels often exhibits overlapping signal characteristics, particularly within the range of mild to moderate muscle activations. This overlap reduces the discriminability of contraction states in the 25%–50% maximum voluntary contraction (MVC) range, making it challenging for iEMG-based prediction systems to achieve reliable classification. As a consequence, such systems frequently produce intermittent or discontinuous outputs, which compromise the naturalness and stability of prosthetic and exoskeleton control. To address this limitation, we propose the L2HiEMG synthesizer, a neural network model built upon the CycleGAN framework. The model is designed to transform low-intensity iEMG signals (25% MVC) into their higher-intensity counterparts (50% MVC), thereby simulating the nonlinear dynamics of motor unit recruitment and muscle fiber activation. Leveraging the dual Generator–Discriminator structure of CycleGAN and its cycle-consistency constraint, the proposed method effectively addresses issues arising from unpaired data, sample imbalance, and ambiguous label correspondence. Experimental evaluations demonstrated that the L2HiEMG synthesizer can generate high-fidelity signals at higher intensity. In the time domain, the synthesized signals exhibited mean and standard deviation differences of less than 0.01 and 0.04, respectively, with correlation coefficients consistently exceeding 0.99. Frequency-domain analyses further validated the accuracy of the generated signals, showing minimal deviations in zero-crossing rate (<span><math><mo><</mo></math></span>0.08), signal energy (<span><math><mo><</mo></math></span>0.01), and power spectral density (<span><math><mo><</mo></math></span>0.03). Taken together, these results confirmed the model’s capability to produce physiologically realistic high-intensity iEMG signals, thereby offering a promising strategy to enhance the stability and responsiveness of iEMG-driven assistive control systems.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109364"},"PeriodicalIF":4.9,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145792053","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

Uncertainty-weighted feature alignment and information interaction network for multi-scale medical image segmentation 多尺度医学图像分割的不确定加权特征对齐与信息交互网络

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-20 DOI: 10.1016/j.bspc.2025.109399

Min Zhang , Junxia Wang , Junkai Wang , Yuanjie Zheng

Lesion segmentation is a critical task in modern clinical applications, where the scale variability of lesions poses significant challenges, making multi-scale feature learning a key technique. However, existing multi-scale approaches often overlook feature misalignment during cross-resolution fusion, imbalanced interaction between hierarchical features, and insufficient utilization of background information, which substantially degrades performance on small lesions and ambiguous boundaries. Moreover, learning target shifts caused by a minority of high-uncertainty features can impair model generalization and boundary prediction stability. To address these challenges, we propose UAINet, a novel network that performs uncertainty-weighted feature alignment and dynamic feature interaction for medical image segmentation. Specifically, the Uncertainty-weighted Offset Alignment Module (UOAM) constructs uncertainty-aware maps via entropy modeling to emphasize reliable features, guiding pixel-level offset learning for precise cross-resolution feature alignment, thereby effectively mitigating detail loss caused by feature misalignment. Moreover, we propose the Context-aware Feature Interaction Module (CAIM) that bridges the gap between local and global features by adaptively modulating their interactions, ensuring balanced information flow and improves multi-level feature consistency. Meanwhile, the Boundary-aware Attention Strategy (BAS) models pixel correlations within the foreground and background regions separately, facilitating collaborative learning between them to fully exploit latent boundary cues, enhancing the model’s ability to handle fuzzy boundaries. We evaluate UAINet on three publicly available medical image segmentation datasets: ISIC2016, ISIC2017, and BUSI. Experimental results show that UAINet achieves state-of-the-art performance, with substantial improvements in Dice score and Intersection over Union (IoU), demonstrating the effectiveness of the proposed method.

病灶分割是现代临床应用中的一项关键任务，病灶的尺度可变性给分割带来很大挑战，多尺度特征学习成为分割的关键技术。然而，现有的多尺度方法往往忽略了跨分辨率融合过程中的特征不匹配、层次特征之间的交互不平衡以及背景信息的利用不足，这大大降低了在小病灶和模糊边界上的性能。此外，由少数高不确定性特征引起的学习目标移位会损害模型的泛化和边界预测的稳定性。为了解决这些挑战，我们提出了UAINet，一种用于医学图像分割的新型网络，它执行不确定性加权特征对齐和动态特征交互。其中，不确定性加权偏移对齐模块（UOAM）通过熵建模构建不确定性感知地图，强调可靠特征，引导像素级偏移学习进行精确的跨分辨率特征对齐，从而有效减轻特征不对齐带来的细节损失。此外，我们提出了上下文感知特征交互模块（CAIM），该模块通过自适应调节局部和全局特征之间的相互作用，弥合了局部和全局特征之间的差距，确保了信息流的平衡，提高了多层次特征的一致性。同时，边界感知注意策略（BAS）分别对前景和背景区域内的像素相关性进行建模，促进它们之间的协同学习，充分利用潜在的边界线索，增强模型处理模糊边界的能力。我们在三个公开可用的医学图像分割数据集（ISIC2016、ISIC2017和BUSI）上评估了UAINet。实验结果表明，UAINet达到了最先进的性能，在骰子得分和交集比联合（IoU）方面有了实质性的改进，证明了所提出方法的有效性。

{"title":"Uncertainty-weighted feature alignment and information interaction network for multi-scale medical image segmentation","authors":"Min Zhang , Junxia Wang , Junkai Wang , Yuanjie Zheng","doi":"10.1016/j.bspc.2025.109399","DOIUrl":"10.1016/j.bspc.2025.109399","url":null,"abstract":"<div><div>Lesion segmentation is a critical task in modern clinical applications, where the scale variability of lesions poses significant challenges, making multi-scale feature learning a key technique. However, existing multi-scale approaches often overlook feature misalignment during cross-resolution fusion, imbalanced interaction between hierarchical features, and insufficient utilization of background information, which substantially degrades performance on small lesions and ambiguous boundaries. Moreover, learning target shifts caused by a minority of high-uncertainty features can impair model generalization and boundary prediction stability. To address these challenges, we propose UAINet, a novel network that performs uncertainty-weighted feature alignment and dynamic feature interaction for medical image segmentation. Specifically, the Uncertainty-weighted Offset Alignment Module (UOAM) constructs uncertainty-aware maps via entropy modeling to emphasize reliable features, guiding pixel-level offset learning for precise cross-resolution feature alignment, thereby effectively mitigating detail loss caused by feature misalignment. Moreover, we propose the Context-aware Feature Interaction Module (CAIM) that bridges the gap between local and global features by adaptively modulating their interactions, ensuring balanced information flow and improves multi-level feature consistency. Meanwhile, the Boundary-aware Attention Strategy (BAS) models pixel correlations within the foreground and background regions separately, facilitating collaborative learning between them to fully exploit latent boundary cues, enhancing the model’s ability to handle fuzzy boundaries. We evaluate UAINet on three publicly available medical image segmentation datasets: ISIC2016, ISIC2017, and BUSI. Experimental results show that UAINet achieves state-of-the-art performance, with substantial improvements in Dice score and Intersection over Union (IoU), demonstrating the effectiveness of the proposed method.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109399"},"PeriodicalIF":4.9,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791835","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 shallow 2D self-attention-CNN framework for depression severity identification using EEG functional connectivity network 基于脑电功能连接网络的浅二维自注意- cnn框架抑郁症严重程度识别

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-20 DOI: 10.1016/j.bspc.2025.109397

Yihan Zhou , Xiaokang Yu , Huiping Lin , Rihui Li , Jiuxing Liang , Xue Shi , Yuxi Luo

Depression inflicts significant harm on both society and family. Previous studies have indicated that the functional network of EEG signals worked well in recognizing major depression. This study aims to further identify depression severities and characterize their EEG network difference by designing a deep learning strategy and a corresponding visualization method. Weighted phase lag indexes were computed across four frequency sub-bands to delineate the functional connectivity (FC), serving as input matrix. To better adapt volume conduction effects, a shallow CNN-based incorporated with 2D Self-Attention architecture was designed, enabling the model to capture information across diverse spans and scales within the FC matrix. Leveraging the Grad-CAM algorithm, the model highlighted crucial FCs and corresponding EEG pairs for classification. Finally, the changes of EEG network across depression severities were statistically analyzed and manifested. Rest state EEG recordings from 30 healthy controls, 35 mild depressed and 26 severe depressed patients were included. An accuracy of 89.2 % was achieved in tri-classification of 10-second EEG segments using 52 EEG channels, 84.1 % with 30 selected channels, which outperformed the classical classifiers. Notably, the results revealed that significant FC changes from mild to severe depression didn’t exhibit a simple or monotonous pattern. This research presented a directly measured methodology to identify depression severity, vital for informing prevention and therapeutic interventions. Furthermore, the findings shed light on the evolving patterns of brain function as depression progresses. The proposed deep learning model and channel selection algorithm offer potential applications beyond this study, promising broader utility in EEG-based research endeavors.

抑郁症对社会和家庭都造成重大伤害。以往的研究表明，脑电图信号的功能网络在识别重度抑郁症方面起着很好的作用。本研究旨在通过设计深度学习策略和相应的可视化方法，进一步识别抑郁症的严重程度，并表征其脑电网络差异。加权相位滞后指数在四个频率子带上计算，以描绘功能连接（FC），作为输入矩阵。为了更好地适应体积传导效应，设计了一种基于浅层cnn的二维自注意结构，使模型能够在FC矩阵中捕获不同跨度和尺度的信息。利用Grad-CAM算法，该模型突出了关键的fc和相应的EEG对进行分类。最后，对不同抑郁程度的脑电网络变化进行统计分析。其中包括30名健康对照者、35名轻度抑郁症患者和26名重度抑郁症患者的静息状态EEG记录。采用52个通道对10秒脑电信号进行三分类，准确率达89.2%，30个通道对10秒脑电信号进行三分类，准确率达84.1%，优于经典分类器。值得注意的是，结果显示，从轻度到重度抑郁的显著FC变化并不表现出简单或单调的模式。这项研究提出了一种直接测量的方法来确定抑郁症的严重程度，这对于告知预防和治疗干预措施至关重要。此外，研究结果还揭示了随着抑郁症的发展，大脑功能的演变模式。所提出的深度学习模型和通道选择算法提供了超出本研究的潜在应用，在基于脑电图的研究工作中有更广泛的应用前景。

{"title":"A shallow 2D self-attention-CNN framework for depression severity identification using EEG functional connectivity network","authors":"Yihan Zhou , Xiaokang Yu , Huiping Lin , Rihui Li , Jiuxing Liang , Xue Shi , Yuxi Luo","doi":"10.1016/j.bspc.2025.109397","DOIUrl":"10.1016/j.bspc.2025.109397","url":null,"abstract":"<div><div>Depression inflicts significant harm on both society and family. Previous studies have indicated that the functional network of EEG signals worked well in recognizing major depression. This study aims to further identify depression severities and characterize their EEG network difference by designing a deep learning strategy and a corresponding visualization method. Weighted phase lag indexes were computed across four frequency sub-bands to delineate the functional connectivity (FC), serving as input matrix. To better adapt volume conduction effects, a shallow CNN-based incorporated with 2D Self-Attention architecture was designed, enabling the model to capture information across diverse spans and scales within the FC matrix. Leveraging the Grad-CAM algorithm, the model highlighted crucial FCs and corresponding EEG pairs for classification. Finally, the changes of EEG network across depression severities were statistically analyzed and manifested. Rest state EEG recordings from 30 healthy controls, 35 mild depressed and 26 severe depressed patients were included. An accuracy of 89.2 % was achieved in tri-classification of 10-second EEG segments using 52 EEG channels, 84.1 % with 30 selected channels, which outperformed the classical classifiers. Notably, the results revealed that significant FC changes from mild to severe depression didn’t exhibit a simple or monotonous pattern. This research presented a directly measured methodology to identify depression severity, vital for informing prevention and therapeutic interventions. Furthermore, the findings shed light on the evolving patterns of brain function as depression progresses. The proposed deep learning model and channel selection algorithm offer potential applications beyond this study, promising broader utility in EEG-based research endeavors.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109397"},"PeriodicalIF":4.9,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791838","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 hybrid model merging convolutional neural network and differential vision transformer for diabetic retinopathy identification 基于卷积神经网络和差分视觉变压器的糖尿病视网膜病变识别混合模型

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-19 DOI: 10.1016/j.bspc.2025.109435

Jianbo Fan , Nan Xiao , Yuekui Zhang , Ruonan Zhai , Yunxia Chu

This research focuses on the vital challenge of classifying diabetic retinopathy, presenting an improved Conformer model that incorporates a differential vision transformer. The designed model aims to capture complex spatial and temporal features in fundus images—elements that are essential for precise disease classification. For verifying the model’s effectiveness, three well-known datasets were utilized; collectively, these datasets cover a wide variety of diabetic retinopathy manifestations. Within the Conformer architecture, the differential vision transformer module facilitates more efficient extraction of fine-grained details, thereby boosting the model’s ability to distinguish between different disease states. Experimental findings show that the enhanced Conformer performs better than conventional classification approaches in three key metrics: accuracy, sensitivity, and specificity. This study not only adds value to the advancement of cutting-edge medical image analysis technologies but also exhibits great potential for enabling early and accurate diagnosis of diabetic retinopathy in real-world clinical environments.

本研究聚焦于糖尿病视网膜病变分类的关键挑战，提出了一种改进的融合了差动视觉变压器的Conformer模型。设计的模型旨在捕获眼底图像中复杂的时空特征，这些特征对于精确的疾病分类至关重要。为了验证模型的有效性，使用了三个知名的数据集；总的来说，这些数据集涵盖了各种各样的糖尿病视网膜病变表现。在Conformer架构中，差分视觉转换器模块有助于更有效地提取细粒度的细节，从而提高模型区分不同疾病状态的能力。实验结果表明，增强的Conformer在三个关键指标上优于传统的分类方法：准确性、灵敏度和特异性。这项研究不仅为尖端医学图像分析技术的发展增加了价值，而且在现实临床环境中显示出早期准确诊断糖尿病视网膜病变的巨大潜力。

{"title":"A hybrid model merging convolutional neural network and differential vision transformer for diabetic retinopathy identification","authors":"Jianbo Fan , Nan Xiao , Yuekui Zhang , Ruonan Zhai , Yunxia Chu","doi":"10.1016/j.bspc.2025.109435","DOIUrl":"10.1016/j.bspc.2025.109435","url":null,"abstract":"<div><div>This research focuses on the vital challenge of classifying diabetic retinopathy, presenting an improved Conformer model that incorporates a differential vision transformer. The designed model aims to capture complex spatial and temporal features in fundus images—elements that are essential for precise disease classification. For verifying the model’s effectiveness, three well-known datasets were utilized; collectively, these datasets cover a wide variety of diabetic retinopathy manifestations. Within the Conformer architecture, the differential vision transformer module facilitates more efficient extraction of fine-grained details, thereby boosting the model’s ability to distinguish between different disease states. Experimental findings show that the enhanced Conformer performs better than conventional classification approaches in three key metrics: accuracy, sensitivity, and specificity. This study not only adds value to the advancement of cutting-edge medical image analysis technologies but also exhibits great potential for enabling early and accurate diagnosis of diabetic retinopathy in real-world clinical environments.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109435"},"PeriodicalIF":4.9,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145792051","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 two-step feature selection framework with L1L2R2 and ensemble hyperparameter tuning for predicting lung cancer: integrating stacked ensemble models 基于L1L2R2和集成超参数调优的两步特征选择框架预测肺癌：整合堆叠集成模型

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-19 DOI: 10.1016/j.bspc.2025.109428

Vahiduddin Shariff , P. Chiranjeevi , A. Krishna Mohan

This work offers a robust strategy for lung cancer prediction by aggregating advanced techniques for feature selection and ensemble learning application. We propose to use a model combining L1 (Lasso) and L2 (Ridge) regularisation methods including L1L2R² regularisation. This model seeks to enhance the prediction characterizations identification. The novelty of this study lies in the integration of a two-step hybrid feature selection mechanism with a stacked ensemble classifier, further enhanced by a three-level Ensemble Hyperparameter Tuning (EHT) approach using Random Search, Grid Search, and Bayesian Optimisation. This work offers a robust strategy for lung cancer prediction by aggregating advanced techniques for feature selection and ensemble learning application. We propose a model combining L1 (Lasso) and L2 (Ridge) regularisation methods via L1L2R² regularisation. The model integrates Gradient Boosting, Random Forest, and XGBoost classifiers. Ensemble Hyperparameter Tuning (EHT), involving Random Search, Grid Search, and Bayesian Optimisation, optimizes the model’s parameters. The proposed model achieved 97.93 % accuracy, 98.13 % precision, 97.76 % recall, 97.94 % F1-score, 99.89% AUC, and a low RMSE of 0.1439, outperforming methods like PSO-RF (92 %), RFE-Light GBM (87.45 %), and TABU-SDS-NN (94.07 %). These results demonstrate the model’s high reliability for accurate lung cancer prediction. The study encourages the adoption of hybrid feature selection and ensemble methods in medical diagnosis and proposes exploring larger datasets and broader clinical applications in future work. This study stresses the use of advanced feature selection and ensemble methods to increase medical diagnosis accuracy. It proposes utilising larger datasets and investigating more therapeutic uses in future study.

这项工作通过聚合先进的特征选择和集成学习应用技术，为肺癌预测提供了一个强大的策略。我们建议使用结合L1 （Lasso）和L2 （Ridge）正则化方法的模型，包括L1L2R2正则化。该模型旨在提高预测特征识别。本研究的新颖之处在于将两步混合特征选择机制与堆叠集成分类器集成在一起，并通过使用随机搜索、网格搜索和贝叶斯优化的三层集成超参数调优（EHT）方法进一步增强。这项工作通过聚合先进的特征选择和集成学习应用技术，为肺癌预测提供了一个强大的策略。我们提出了一个结合L1 （Lasso）和L2 （Ridge）正则化方法的L1L2R2正则化模型。该模型集成了梯度增强、随机森林和XGBoost分类器。集成超参数调整（EHT），包括随机搜索，网格搜索和贝叶斯优化，优化模型的参数。该模型的准确率为97.93%，精密度为98.13%，召回率为97.76%，f1评分为97.94%，AUC为99.89%，RMSE为0.1439，优于PSO-RF (92%), RFE-Light GBM（87.45%）和TABU-SDS-NN（94.07%）等方法。这些结果表明该模型对肺癌的准确预测具有很高的可靠性。该研究鼓励在医学诊断中采用混合特征选择和集成方法，并提出在未来的工作中探索更大的数据集和更广泛的临床应用。本研究强调使用先进的特征选择和集成方法来提高医学诊断的准确性。它建议在未来的研究中利用更大的数据集并调查更多的治疗用途。

{"title":"A two-step feature selection framework with L1L2R2 and ensemble hyperparameter tuning for predicting lung cancer: integrating stacked ensemble models","authors":"Vahiduddin Shariff , P. Chiranjeevi , A. Krishna Mohan","doi":"10.1016/j.bspc.2025.109428","DOIUrl":"10.1016/j.bspc.2025.109428","url":null,"abstract":"<div><div>This work offers a robust strategy for lung cancer prediction by aggregating advanced techniques for feature selection and ensemble learning application. We propose to use a model combining L1 (Lasso) and L2 (Ridge) regularisation methods including L1L2R<sup>2</sup> regularisation. This model seeks to enhance the prediction characterizations identification. The novelty of this study lies in the integration of a two-step hybrid feature selection mechanism with a stacked ensemble classifier, further enhanced by a three-level Ensemble Hyperparameter Tuning (EHT) approach using Random Search, Grid Search, and Bayesian Optimisation. This work offers a robust strategy for lung cancer prediction by aggregating advanced techniques for feature selection and ensemble learning application. We propose a model combining L1 (Lasso) and L2 (Ridge) regularisation methods via L1L2R<sup>2</sup> regularisation. The model integrates Gradient Boosting, Random Forest, and XGBoost classifiers. Ensemble Hyperparameter Tuning (EHT), involving Random Search, Grid Search, and Bayesian Optimisation, optimizes the model’s parameters. The proposed model achieved 97.93% accuracy, 98.13% precision, 97.76% recall, 97.94% F1-score, 99.89% AUC, and a low RMSE of 0.1439, outperforming methods like PSO-RF (92%), RFE-Light GBM (87.45%), and TABU-SDS-NN (94.07%). These results demonstrate the model’s high reliability for accurate lung cancer prediction. The study encourages the adoption of hybrid feature selection and ensemble methods in medical diagnosis and proposes exploring larger datasets and broader clinical applications in future work. This study stresses the use of advanced feature selection and ensemble methods to increase medical diagnosis accuracy. It proposes utilising larger datasets and investigating more therapeutic uses in future study.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109428"},"PeriodicalIF":4.9,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791843","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

Frequency-spatial domain feature interaction for accurate cell detection in multiplex immunohistochemistry images 在多重免疫组织化学图像中精确检测细胞的频率-空间域特征相互作用

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-19 DOI: 10.1016/j.bspc.2025.109392

Xinwen Zhou , Ke Cheng , Jingyuan Yang , Ruiping Dai , Ran Wei , Jingting Jiang

Multiplex Immunohistochemistry (mIHC) simultaneously maps multiple cell-type biomarkers within a single tissue section, revealing the spatial architecture and cellular interactions of the tumor microenvironment (TME). This co-localization capability is pivotal for advancing clinical diagnostics and therapeutics. Consequently, accurate cell detection and classification in mIHC images are fundamental to deciphering TME dynamics. However, due to issues such as heterogeneous staining intensities, intricate cellular textures, and cell adhesion artifacts in mIHC images, cell detection is prone to background interference and blurred cell boundaries, posing significant challenges for existing methods. This paper proposes an end-to-end point-based framework, the Multi-level Feature Enhancement Network (MFE-Net), to suppress image background interference and enhance cell boundaries. First, to mitigate the impact of staining heterogeneity and complex textures, we introduce a Frequency-Spatial Interaction (FSI) module, comprising a frequency-domain branch with dynamic filters for global texture modeling and a spatial-domain branch for local structural refinement. Second, an Upsampling Edge Enhancement (UEE) module is designed to preserve boundary details during resolution recovery. Third, a Mixed Query Selection (MQS) strategy enhances detection robustness in complex backgrounds. Evaluated on a curated mIHC dataset, MFE-Net achieves state-of-the-art performance, with cell detection and classification F1 scores of 0.795 and 0.780, respectively, outperforming existing methods.

多重免疫组织化学（Multiplex Immunohistochemistry, mIHC）在单个组织切片中同时绘制多种细胞型生物标志物，揭示肿瘤微环境（tumor microenvironment， TME）的空间结构和细胞相互作用。这种共定位能力对于推进临床诊断和治疗至关重要。因此，在mIHC图像中准确的细胞检测和分类是破译TME动力学的基础。然而，由于mIHC图像中的染色强度不均、细胞纹理复杂和细胞粘附伪影等问题，细胞检测容易受到背景干扰和细胞边界模糊，对现有方法提出了重大挑战。本文提出了一种端到端的基于点的框架——多级特征增强网络（MFE-Net）来抑制图像背景干扰和增强细胞边界。首先，为了减轻染色异质性和复杂纹理的影响，我们引入了一个频率-空间相互作用（FSI）模块，该模块包括一个带有动态滤波器的频率域分支，用于全局纹理建模，以及一个用于局部结构细化的空间域分支。其次，设计了上采样边缘增强（UEE）模块，以在分辨率恢复期间保留边界细节。第三，混合查询选择（MQS）策略增强了复杂背景下的检测鲁棒性。在精心设计的mIHC数据集上进行评估，MFE-Net达到了最先进的性能，细胞检测和分类F1得分分别为0.795和0.780，优于现有方法。

{"title":"Frequency-spatial domain feature interaction for accurate cell detection in multiplex immunohistochemistry images","authors":"Xinwen Zhou , Ke Cheng , Jingyuan Yang , Ruiping Dai , Ran Wei , Jingting Jiang","doi":"10.1016/j.bspc.2025.109392","DOIUrl":"10.1016/j.bspc.2025.109392","url":null,"abstract":"<div><div>Multiplex Immunohistochemistry (mIHC) simultaneously maps multiple cell-type biomarkers within a single tissue section, revealing the spatial architecture and cellular interactions of the tumor microenvironment (TME). This co-localization capability is pivotal for advancing clinical diagnostics and therapeutics. Consequently, accurate cell detection and classification in mIHC images are fundamental to deciphering TME dynamics. However, due to issues such as heterogeneous staining intensities, intricate cellular textures, and cell adhesion artifacts in mIHC images, cell detection is prone to background interference and blurred cell boundaries, posing significant challenges for existing methods. This paper proposes an end-to-end point-based framework, the Multi-level Feature Enhancement Network (MFE-Net), to suppress image background interference and enhance cell boundaries. First, to mitigate the impact of staining heterogeneity and complex textures, we introduce a Frequency-Spatial Interaction (FSI) module, comprising a frequency-domain branch with dynamic filters for global texture modeling and a spatial-domain branch for local structural refinement. Second, an Upsampling Edge Enhancement (UEE) module is designed to preserve boundary details during resolution recovery. Third, a Mixed Query Selection (MQS) strategy enhances detection robustness in complex backgrounds. Evaluated on a curated mIHC dataset, MFE-Net achieves state-of-the-art performance, with cell detection and classification F1 scores of 0.795 and 0.780, respectively, outperforming existing methods.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109392"},"PeriodicalIF":4.9,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791888","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

APP-Net: A physics-constrained joint prior-guided network for CBCT metal artifact reduction APP-Net：一个物理约束的联合先验引导网络，用于CBCT金属伪影还原

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-19 DOI: 10.1016/j.bspc.2025.109433

Kai Chen , Zihao Wang , Shipeng Xie , Hui Tang , Zhan Wu , Chunfeng Yang , Shangwen Yang , Jiabing Gu , Yan Xi , Yang Chen

As one of the commonly available instruments in hospitals, Cone-beam computed tomography (CBCT), is widely used in clinical tasks such as disease diagnosis, interventional surgery, and radiation therapy due to its high spatial resolution, flexibility, and high X-ray utilization. In some practical applications, metallic implants within the patient’s body cause the CBCT projection to be contaminated, and the reconstructed CBCT image suffers from shadowing, which is detrimental to the clinical diagnosis. Therefore, developing high-performance CBCT metal artifact reduction (MAR) algorithms for CBCT reconstruction and clinical diagnosis is imperative. Existing MAR methods for CBCT fail to balance the accuracy of restoring tissue details around metals with artifact removal in metal-free regions in CBCT images. In this paper, we proposed a Physics-constrained joint Prior-guided Network (APP-Net) for CBCT metal artifact reduction, which efficiently combines physical constraints for beam hardening correction in the projection domain and a data-driven self-prior boosted restoration module in the image domain for MAR in CBCT. Qualitative and quantitative results on simulated and public datasets demonstrate the superior performance of APP-Net. Additionally, a preliminary feasibility study on clinical data indicates the potential transferability of the method to real-world scenarios.

锥形束计算机断层扫描（Cone-beam computed tomography， CBCT）是医院常用的仪器之一，由于其高空间分辨率、灵活性和高x线利用率，被广泛应用于疾病诊断、介入手术、放射治疗等临床任务。在一些实际应用中，患者体内的金属植入物会导致CBCT投影受到污染，重建的CBCT图像会出现阴影，不利于临床诊断。因此，开发高性能的CBCT金属伪影还原（MAR）算法用于CBCT重建和临床诊断势在必行。现有的CBCT MAR方法无法平衡CBCT图像中金属周围组织细节的恢复和无金属区域伪影的去除的准确性。本文提出了一种基于物理约束的联合先验引导网络（APP-Net）用于CBCT金属伪影还原，该网络有效地结合了用于投影域光束硬化校正的物理约束和用于CBCT中MAR的图像域数据驱动的自先验增强恢复模块。在模拟和公共数据集上的定性和定量结果证明了APP-Net的优越性能。此外，对临床数据的初步可行性研究表明，该方法在现实世界中具有潜在的可移植性。

{"title":"APP-Net: A physics-constrained joint prior-guided network for CBCT metal artifact reduction","authors":"Kai Chen , Zihao Wang , Shipeng Xie , Hui Tang , Zhan Wu , Chunfeng Yang , Shangwen Yang , Jiabing Gu , Yan Xi , Yang Chen","doi":"10.1016/j.bspc.2025.109433","DOIUrl":"10.1016/j.bspc.2025.109433","url":null,"abstract":"<div><div>As one of the commonly available instruments in hospitals, Cone-beam computed tomography (CBCT), is widely used in clinical tasks such as disease diagnosis, interventional surgery, and radiation therapy due to its high spatial resolution, flexibility, and high X-ray utilization. In some practical applications, metallic implants within the patient’s body cause the CBCT projection to be contaminated, and the reconstructed CBCT image suffers from shadowing, which is detrimental to the clinical diagnosis. Therefore, developing high-performance CBCT metal artifact reduction (MAR) algorithms for CBCT reconstruction and clinical diagnosis is imperative. Existing MAR methods for CBCT fail to balance the accuracy of restoring tissue details around metals with artifact removal in metal-free regions in CBCT images. In this paper, we proposed a Physics-constrained joint Prior-guided Network (APP-Net) for CBCT metal artifact reduction, which efficiently combines physical constraints for beam hardening correction in the projection domain and a data-driven self-prior boosted restoration module in the image domain for MAR in CBCT. Qualitative and quantitative results on simulated and public datasets demonstrate the superior performance of APP-Net. Additionally, a preliminary feasibility study on clinical data indicates the potential transferability of the method to real-world scenarios.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109433"},"PeriodicalIF":4.9,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791847","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

Fetal brain anomaly identification from ultrasound images using multiclass classifiers 利用多分类器识别超声胎儿脑异常

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-19 DOI: 10.1016/j.bspc.2025.109382

Rehan Ahmad

Detection of fetal brain anomaly at an early stage can save, baby from life-long issues, parents from emotional, financial and societal burden. It helps clinician to decide course of treatment and medical procedure. The earlier works in this domain focused on identification of normal and abnormal class. When fetal brain images of multiple diseases considered, had low classification accuracy. In this work, a multiclass classification scheme for identification of fetal brain anomalies from open access dataset used. As per discussions and suggestion from clinician fetal brain images of arachnoid-cyst, anold-chiari-malformation, cerebellah-hypoplasia, colphocephaly and normal fetal brain ultrasound images used in this work. Gray level co-occurrence matrix (GLCM) features extracted from the images belonging to above mentioned fetal brain anomalies. These GLCM feature used to train and test multiclass classifiers such as linear discriminant analysis (LDA), support vector machine (one versus one and one versus all) with linear kernel, light weight neural networks like Narrow, Medium, Trilayered and Optimized neural networks. From simulation results, it has been observed that, LDA multiclass classifier outperforms all the other classifiers considered in this work. LDA identifies normal, arachnoid cyst and cerebellah hypoplasia images with 100% accuracy and anold-chiari-malformation, colphocephaly with 89% and 93% accuracy respectively. Also, when the results of proposed work compared with similar works, it is observed that, LDA identifies normal images with 19% increased accuracy and cerebellah hypoplasia and colphocephaly anomaly with an increment of 8% accuracy respectively.

早期发现胎儿脑异常可以使婴儿免于终身问题，使父母免于情感、经济和社会负担。它有助于临床医生决定治疗过程和医疗程序。该领域的早期工作主要集中在正常类和异常类的识别上。当考虑多种疾病的胎儿脑图像时，分类准确率较低。在这项工作中，使用了一种多类分类方案来识别开放获取数据集中的胎儿脑异常。根据本研究中使用的蛛网膜囊肿、老年性睫状体畸形、小脑发育不全、colphocephaly和正常胎儿脑超声图像的临床医生讨论和建议。从上述胎儿脑异常图像中提取灰度共生矩阵（GLCM）特征。这些GLCM特征用于训练和测试多类分类器，如线性判别分析（LDA），支持向量机（一对一和一对一）与线性核，轻权重神经网络，如Narrow， Medium， Trilayered和Optimized神经网络。从仿真结果可以看出，LDA多类分类器优于本研究中考虑的所有其他分类器。LDA对正常、蛛网膜囊肿和小脑发育不全图像的识别准确率为100%，对老年脊膜畸形、小脑畸形的识别准确率分别为89%和93%。与同类研究结果比较，LDA对正常图像的识别准确率提高了19%，对小脑发育不全和colphocephaly异常的识别准确率分别提高了8%。

{"title":"Fetal brain anomaly identification from ultrasound images using multiclass classifiers","authors":"Rehan Ahmad","doi":"10.1016/j.bspc.2025.109382","DOIUrl":"10.1016/j.bspc.2025.109382","url":null,"abstract":"<div><div>Detection of fetal brain anomaly at an early stage can save, baby from life-long issues, parents from emotional, financial and societal burden. It helps clinician to decide course of treatment and medical procedure. The earlier works in this domain focused on identification of normal and abnormal class. When fetal brain images of multiple diseases considered, had low classification accuracy. In this work, a multiclass classification scheme for identification of fetal brain anomalies from open access dataset used. As per discussions and suggestion from clinician fetal brain images of arachnoid-cyst, anold-chiari-malformation, cerebellah-hypoplasia, colphocephaly and normal fetal brain ultrasound images used in this work. Gray level co-occurrence matrix (GLCM) features extracted from the images belonging to above mentioned fetal brain anomalies. These GLCM feature used to train and test multiclass classifiers such as linear discriminant analysis (LDA), support vector machine (one versus one and one versus all) with linear kernel, light weight neural networks like Narrow, Medium, Trilayered and Optimized neural networks. From simulation results, it has been observed that, LDA multiclass classifier outperforms all the other classifiers considered in this work. LDA identifies normal, arachnoid cyst and cerebellah hypoplasia images with 100% accuracy and anold-chiari-malformation, colphocephaly with 89% and 93% accuracy respectively. Also, when the results of proposed work compared with similar works, it is observed that, LDA identifies normal images with 19% increased accuracy and cerebellah hypoplasia and colphocephaly anomaly with an increment of 8% accuracy respectively.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109382"},"PeriodicalIF":4.9,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791833","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

Machine Learning-assisted Quorum Sensing Monitoring and Control Systems for Precision Gene Regulation: Revolutionizing Synthetic Biology and Autonomous Therapeutic Applications 精确基因调控的机器学习辅助群体感应监测和控制系统：革命性的合成生物学和自主治疗应用

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-19 DOI: 10.1016/j.bspc.2025.109285

Dang Anh Tuan , Pham Vu Nhat Uyen

Quorum sensing (QS) coordinates population-level microbial behaviors, but conventional circuits are largely static and brittle to changing conditions. This narrative review synthesizes how machine learning (ML) elevates QS from open-loop regulation to adaptive, closed-loop control across biomanufacturing, environmental, and precision-medicine contexts. Surveying peer-reviewed work (2010–Aug 2025) across major biomedical and engineering databases, we distill three consistent gains: robust pattern recognition from noisy QS signals, predictive scheduling of gene-expression programs, and real-time feedback control that maintains performance under drift and perturbation. Reported exemplars—evaluated with time-aware protocols—include up to ∼ 45% biofilm reduction with ∼ 30% antibiotic potentiation in Pseudomonas aeruginosa, ∼60% toxin reduction in pathogenic E. coli, ∼70% cholera-toxin reduction in Vibrio cholerae, and ∼ 50% toxin reduction with biofilm prevention in Staphylococcus aureus. We compare controller families (PID/MPC/RL), emphasize mechanism-aware (gray-box) models and stable explanations (e.g., SHAP/IG/TCAV) to align ML decisions with QS biology, and outline translational guardrails—predictive monitoring, redundancy, kill-switches, human-in-the-loop oversight, and auditable change control—to manage failure modes and distributional shift. Practical recommendations include standardized units/normalization, time-aware cross-validation with explicit baselines, and deployment-oriented evaluation (latency budgets, fault-injection, fallback behavior). Quantitative statements are mapped to primary sources for auditability in Supplementary Table S0. Collectively, the evidence indicates that ML-enabled QS can deliver material performance gains when paired with transparent validation and layered safety, while highlighting priorities for open benchmarks and prospective, real-world studies. No new experiments or datasets are reported.

群体感应（Quorum sensing， QS）协调群体水平的微生物行为，但传统的电路在很大程度上是静态的，易受环境变化的影响。这篇叙述性综述综合了机器学习（ML）如何将QS从开环调节提升到生物制造、环境和精密医学环境中的自适应闭环控制。通过对主要生物医学和工程数据库的同行评审工作（2010年8月- 2025年8月）的调查，我们得出了三个一致的成果：从噪声QS信号中稳健的模式识别，基因表达程序的预测调度，以及在漂移和扰动下保持性能的实时反馈控制。已报道的范例——采用时间敏感方案进行评估——包括铜绿假单胞菌中高达45%的生物膜减少与约30%的抗生素增强，致病性大肠杆菌中约60%的毒素减少，霍乱弧菌中约70%的霍乱毒素减少，以及金黄色葡萄球菌中约50%的毒素减少与生物膜预防。我们比较了控制器家族（PID/MPC/RL），强调机制感知（灰盒）模型和稳定的解释（例如，SHAP/IG/TCAV），以使ML决策与QS生物学保持一致，并概述了翻译围栏-预测监测，冗余，死亡开关，人在环监督和可审计的变更控制-以管理故障模式和分布转移。实用的建议包括标准化单元/规范化、具有显式基线的时间感知交叉验证，以及面向部署的评估（延迟预算、故障注入、回退行为）。在补充表50中，定量报表被映射到可审计性的主要来源。总的来说，有证据表明，当与透明验证和分层安全性相结合时，支持ml的QS可以提供材料性能提升，同时突出了开放基准和前瞻性现实世界研究的优先事项。没有新的实验或数据集报告。

{"title":"Machine Learning-assisted Quorum Sensing Monitoring and Control Systems for Precision Gene Regulation: Revolutionizing Synthetic Biology and Autonomous Therapeutic Applications","authors":"Dang Anh Tuan , Pham Vu Nhat Uyen","doi":"10.1016/j.bspc.2025.109285","DOIUrl":"10.1016/j.bspc.2025.109285","url":null,"abstract":"<div><div>Quorum sensing (QS) coordinates population-level microbial behaviors, but conventional circuits are largely static and brittle to changing conditions. This narrative review synthesizes how machine learning (ML) elevates QS from open-loop regulation to adaptive, closed-loop control across biomanufacturing, environmental, and precision-medicine contexts. Surveying peer-reviewed work (2010–Aug 2025) across major biomedical and engineering databases, we distill three consistent gains: robust pattern recognition from noisy QS signals, predictive scheduling of gene-expression programs, and real-time feedback control that maintains performance under drift and perturbation. Reported exemplars—evaluated with time-aware protocols—include up to ∼ 45% biofilm reduction with ∼ 30% antibiotic potentiation in <em>Pseudomonas aeruginosa</em>, ∼60% toxin reduction in pathogenic <em>E. coli</em>, ∼70% cholera-toxin reduction in <em>Vibrio cholerae</em>, and ∼ 50% toxin reduction with biofilm prevention in <em>Staphylococcus aureus</em>. We compare controller families (PID/MPC/RL), emphasize mechanism-aware (gray-box) models and stable explanations (e.g., SHAP/IG/TCAV) to align ML decisions with QS biology, and outline translational guardrails—predictive monitoring, redundancy, kill-switches, human-in-the-loop oversight, and auditable change control—to manage failure modes and distributional shift. Practical recommendations include standardized units/normalization, time-aware cross-validation with explicit baselines, and deployment-oriented evaluation (latency budgets, fault-injection, fallback behavior). Quantitative statements are mapped to primary sources for auditability in <span><span>Supplementary Table S0</span></span>. Collectively, the evidence indicates that ML-enabled QS can deliver material performance gains when paired with transparent validation and layered safety, while highlighting priorities for open benchmarks and prospective, real-world studies. No new experiments or datasets are reported.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109285"},"PeriodicalIF":4.9,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791845","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

EA-LGS: Edge-aware vessel segmentation based on local-to-global supervision EA-LGS：基于局部到全局监督的边缘感知船舶分割

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2025-12-18 DOI: 10.1016/j.bspc.2025.109411

Zizheng Li , Huadeng Wang , Wen Shi , Bo Li , Xiaonan Luo

Vessel segmentation is crucial for the early diagnosis and monitoring of eye diseases, but issues such as low contrast, blurred boundaries, and noise interference often reduce segmentation accuracy. To address these challenges, we propose an edge-aware vessel segmentation method based on local-to-global supervision. This approach integrates frequency-domain enhancement and multi-scale modeling to accurately capture vessel edge information and dynamically optimize the weight of features at different scales. By overcoming the bottleneck of high-frequency information loss in traditional convolutional neural networks, our method significantly improves the segmentation accuracy of small vessels and junctions. Additionally, by combining local and global feature distribution loss, the method enhances detail recovery and maintains global structural consistency. Experimental results show that the segmentation accuracy on the DRIVE, CHASE_DB1, and DCA1 datasets are 97.12%, 97.71%, and 97.82%, respectively. Furthermore, the segmentation results on lesion images further demonstrate the robustness of the method. Source code is available at https://github.com/Lrsm-sudo/EA-LGS.

血管分割对于眼部疾病的早期诊断和监测至关重要，但对比度低、边界模糊和噪声干扰等问题往往会降低分割的准确性。为了解决这些挑战，我们提出了一种基于局部到全局监督的边缘感知船舶分割方法。该方法将频域增强和多尺度建模相结合，准确捕获血管边缘信息，并在不同尺度下动态优化特征权重。该方法克服了传统卷积神经网络高频信息丢失的瓶颈，显著提高了小血管和连接的分割精度。此外，该方法结合局部和全局特征分布损失，增强了细节恢复能力，并保持了全局结构的一致性。实验结果表明，在DRIVE、CHASE_DB1和DCA1数据集上的分割准确率分别为97.12%、97.71%和97.82%。此外，对病变图像的分割结果进一步证明了该方法的鲁棒性。源代码可从https://github.com/Lrsm-sudo/EA-LGS获得。

{"title":"EA-LGS: Edge-aware vessel segmentation based on local-to-global supervision","authors":"Zizheng Li , Huadeng Wang , Wen Shi , Bo Li , Xiaonan Luo","doi":"10.1016/j.bspc.2025.109411","DOIUrl":"10.1016/j.bspc.2025.109411","url":null,"abstract":"<div><div>Vessel segmentation is crucial for the early diagnosis and monitoring of eye diseases, but issues such as low contrast, blurred boundaries, and noise interference often reduce segmentation accuracy. To address these challenges, we propose an edge-aware vessel segmentation method based on local-to-global supervision. This approach integrates frequency-domain enhancement and multi-scale modeling to accurately capture vessel edge information and dynamically optimize the weight of features at different scales. By overcoming the bottleneck of high-frequency information loss in traditional convolutional neural networks, our method significantly improves the segmentation accuracy of small vessels and junctions. Additionally, by combining local and global feature distribution loss, the method enhances detail recovery and maintains global structural consistency. Experimental results show that the segmentation accuracy on the DRIVE, CHASE_DB1, and DCA1 datasets are 97.12%, 97.71%, and 97.82%, respectively. Furthermore, the segmentation results on lesion images further demonstrate the robustness of the method. Source code is available at <span><span>https://github.com/Lrsm-sudo/EA-LGS</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"115 ","pages":"Article 109411"},"PeriodicalIF":4.9,"publicationDate":"2025-12-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791971","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