Biomedical Signal Processing and Control最新文献_第9页

Deep learning-based multimodal fusion of imaging, pathology, and CTCs for early diagnosis of pediatric distal femur osteosarcoma 基于深度学习的影像、病理和ctc多模态融合在小儿股骨远端骨肉瘤早期诊断中的应用

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109558

Dongjian Song , Meng Su , Qiuliang Liu , Da Zhang , Zechen Yan , Qian Zhang , Qi Wang , Hui Zhang , Longyan Shi , Yingzhong Fan , Heying Yang

This study proposes a novel deep learning (DL)-based multimodal diagnostic framework that integrates magnetic resonance imaging (MRI), computed tomography (CT), histopathological slides, and circulating tumor cells (CTCs) data for early and accurate diagnosis of distal femur osteosarcoma (OS) in pediatric patients. Public datasets including The Cancer Imaging Archive (TCIA), The Cancer Genome Atlas (TCGA), and the Gene Expression Omnibus (GEO) provided imaging and genomic data. Preprocessing involved denoising, normalization, slice alignment, and color standardization using Fiji/ImageJ. Pathological features were extracted via transfer learning using pretrained convolutional neural networks (CNNs) like VGG16 and ResNet50. CTCs were detected and classified using flow cytometry, Hough transform, and support vector machine (SVM) algorithms. A multimodal DL architecture was constructed by fusing image, pathology, and CTC feature vectors, and performance was evaluated through cross-validation. The model achieved an accuracy of 92.5%, sensitivity of 88.7%, specificity of 94.3%, and AUC of 0.96 on an independent test set. Incorporating CTC data notably improved performance in metastasis assessment and diagnosis where imaging was inconclusive. The proposed DL-based multimodal model significantly enhances the early diagnostic capacity for pediatric distal femur OS. Its robustness, diagnostic accuracy, and potential for clinical translation make it a promising tool for personalized treatment strategies.

本研究提出了一种新的基于深度学习（DL）的多模式诊断框架，该框架整合了磁共振成像（MRI）、计算机断层扫描（CT）、组织病理学切片和循环肿瘤细胞（ctc）数据，用于儿科患者股骨远端骨肉瘤（OS）的早期准确诊断。包括癌症成像档案（TCIA）、癌症基因组图谱（TCGA）和基因表达综合数据库（GEO）在内的公共数据集提供了成像和基因组数据。预处理包括使用Fiji/ImageJ去噪、归一化、切片对齐和颜色标准化。使用VGG16和ResNet50等预训练卷积神经网络（cnn）通过迁移学习提取病理特征。采用流式细胞术、霍夫变换和支持向量机（SVM）算法对ctc进行检测和分类。通过融合图像、病理和CTC特征向量构建多模态深度学习架构，并通过交叉验证评估性能。该模型在独立测试集上的准确率为92.5%，灵敏度为88.7%，特异性为94.3%，AUC为0.96。合并CTC数据显著提高了在影像学不确定的情况下转移评估和诊断的表现。所提出的基于dl的多模态模型显著提高了小儿股骨远端OS的早期诊断能力。它的稳健性、诊断准确性和临床翻译的潜力使其成为个性化治疗策略的有前途的工具。

{"title":"Deep learning-based multimodal fusion of imaging, pathology, and CTCs for early diagnosis of pediatric distal femur osteosarcoma","authors":"Dongjian Song , Meng Su , Qiuliang Liu , Da Zhang , Zechen Yan , Qian Zhang , Qi Wang , Hui Zhang , Longyan Shi , Yingzhong Fan , Heying Yang","doi":"10.1016/j.bspc.2026.109558","DOIUrl":"10.1016/j.bspc.2026.109558","url":null,"abstract":"<div><div>This study proposes a novel deep learning (DL)-based multimodal diagnostic framework that integrates magnetic resonance imaging (MRI), computed tomography (CT), histopathological slides, and circulating tumor cells (CTCs) data for early and accurate diagnosis of distal femur osteosarcoma (OS) in pediatric patients. Public datasets including The Cancer Imaging Archive (TCIA), The Cancer Genome Atlas (TCGA), and the Gene Expression Omnibus (GEO) provided imaging and genomic data. Preprocessing involved denoising, normalization, slice alignment, and color standardization using Fiji/ImageJ. Pathological features were extracted via transfer learning using pretrained convolutional neural networks (CNNs) like VGG16 and ResNet50. CTCs were detected and classified using flow cytometry, Hough transform, and support vector machine (SVM) algorithms. A multimodal DL architecture was constructed by fusing image, pathology, and CTC feature vectors, and performance was evaluated through cross-validation. The model achieved an accuracy of 92.5%, sensitivity of 88.7%, specificity of 94.3%, and AUC of 0.96 on an independent test set. Incorporating CTC data notably improved performance in metastasis assessment and diagnosis where imaging was inconclusive. The proposed DL-based multimodal model significantly enhances the early diagnostic capacity for pediatric distal femur OS. Its robustness, diagnostic accuracy, and potential for clinical translation make it a promising tool for personalized treatment strategies.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109558"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174416","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

DDMGCN: Deep Dynamic Multi-Graph Convolutional Neural Network for EEG emotion recognition 基于深度动态多图卷积神经网络的脑电情感识别

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109740

Jiao Wang, Zhifen Guo, Peng Zhang, Hongchen Luo, Fengbin Ma, Pengcheng Song

Electroencephalogram (EEG) provides an objective and accurate reflection of the human emotional state, making EEG-based emotion recognition a research focus in fields such as medical measurement and health monitoring. Given the irregular structure of EEG data, graph convolutional neural networks (GCNNs) are effective at learning topological relationships among EEG channels. However, existing GCNN-based work is limited by restricted stacking depth and insufficient flexibility in modeling topological relationships, which makes it difficult to capture complex correlations among EEG signals and ultimately hinders recognition performance. To address these issues, we propose a Deep Dynamic Multi-Graph Convolutional Neural Network (DDMGCN). Specifically, DDMGCN employs a dual-branch collaborative training framework. The master training network integrates a multi-layer 3D convolutional neural network (3DCNN) within the DGCNN architecture to deepen the model, capturing dynamic interactions and multi-level spatiotemporal information. The auxiliary update network introduces a multi-graph structure that adaptively adjusts each layer to achieve optimal topological relationships. Finally, an update strategy leveraging a branch attention mechanism is applied to both branches to optimize model parameters. We evaluate the performance of the DDMGCN on two public SEED and DREAMER datasets. Experimental results, including subject-dependent and subject-independent validations, outperform current state-of-the-art models. This demonstrates the potential of our method for modeling dynamic EEG connectivity in emotion recognition.

脑电图（EEG）能够客观准确地反映人类的情绪状态，使得基于脑电图的情绪识别成为医学测量和健康监测等领域的研究热点。考虑到脑电数据的不规则结构，图卷积神经网络（GCNNs）可以有效地学习脑电通道之间的拓扑关系。然而，现有的基于gcnn的工作受到叠加深度的限制和拓扑关系建模灵活性不足的限制，难以捕获脑电信号之间复杂的相关性，最终影响识别性能。为了解决这些问题，我们提出了一个深度动态多图卷积神经网络（DDMGCN）。具体来说，DDMGCN采用了双分支协作培训框架。主训练网络在DGCNN架构内集成了多层3D卷积神经网络（3DCNN）来深化模型，捕捉动态交互和多层次时空信息。辅助更新网络引入多图结构，自适应调整每一层以实现最优拓扑关系。最后，利用分支关注机制对两个分支应用更新策略来优化模型参数。我们评估了DDMGCN在两个公共SEED和dream数据集上的性能。实验结果，包括主体依赖和主体独立验证，优于当前最先进的模型。这证明了我们的方法在情感识别中建模动态脑电图连接的潜力。

{"title":"DDMGCN: Deep Dynamic Multi-Graph Convolutional Neural Network for EEG emotion recognition","authors":"Jiao Wang, Zhifen Guo, Peng Zhang, Hongchen Luo, Fengbin Ma, Pengcheng Song","doi":"10.1016/j.bspc.2026.109740","DOIUrl":"10.1016/j.bspc.2026.109740","url":null,"abstract":"<div><div>Electroencephalogram (EEG) provides an objective and accurate reflection of the human emotional state, making EEG-based emotion recognition a research focus in fields such as medical measurement and health monitoring. Given the irregular structure of EEG data, graph convolutional neural networks (GCNNs) are effective at learning topological relationships among EEG channels. However, existing GCNN-based work is limited by restricted stacking depth and insufficient flexibility in modeling topological relationships, which makes it difficult to capture complex correlations among EEG signals and ultimately hinders recognition performance. To address these issues, we propose a Deep Dynamic Multi-Graph Convolutional Neural Network (DDMGCN). Specifically, DDMGCN employs a dual-branch collaborative training framework. The master training network integrates a multi-layer 3D convolutional neural network (3DCNN) within the DGCNN architecture to deepen the model, capturing dynamic interactions and multi-level spatiotemporal information. The auxiliary update network introduces a multi-graph structure that adaptively adjusts each layer to achieve optimal topological relationships. Finally, an update strategy leveraging a branch attention mechanism is applied to both branches to optimize model parameters. We evaluate the performance of the DDMGCN on two public SEED and DREAMER datasets. Experimental results, including subject-dependent and subject-independent validations, outperform current state-of-the-art models. This demonstrates the potential of our method for modeling dynamic EEG connectivity in emotion recognition.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109740"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174642","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

Box-guided class-contextual representation learning for self-visual lung nodule detection 盒引导类-上下文表征学习用于自视肺结节检测

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109731

Kefu Zhao, Lei Zhang, Le Yi, Xiuyuan Xu

Detecting lung nodules from computed tomography images has become the routine practice for early lung cancer screening. Conventional methods have made progress in discriminating the high similarity between nodules and non-nodules by extracting contextual information through attention-based or multi-scale mechanisms. However, these methods fail to guide models to focus on learning robust contextual representations in contextual regions as they lack explicit mechanisms to perceive these ambiguous regions, leading to a suboptimal discriminability for such similarity. Moreover, by mimicking clinicians’ diagnostic processes, providing visual evidence of learned regions can enhance the interpretability of detection tools and guide models to generate reliable decisions. In this paper, we propose class-contextual representation learning with ante-hoc visual interpretations to enhance lung nodule detection. It leverages a query-based mechanism via class-contextual vectors to explicitly perceive ambiguous contextual regions for guiding the learning of contextual representations. According to the manifold assumption, this query compresses the contextual information embedded in the feature manifold into a low-dimensional, class-contextual latent space, thereby filtering redundancy for learning discriminative representations. To ensure the efficacy of this query, we propose a box-guided instance-level method that enables class-contextual vectors align with discriminative representations of their respective classes. Our method exhibits a reliable intrinsic visualization effect, which enhances the transparency of the model’s decision-making process and provides learning guidance for further performance gains. Experiments demonstrate that this method achieves competition performance metrics of 93.38% and 65.58% on the LUNA16 and PN9 datasets, respectively.

从计算机断层图像中检测肺结节已成为早期肺癌筛查的常规做法。传统方法通过基于注意或多尺度的机制提取上下文信息，在判别结节与非结节的高度相似性方面取得了进展。然而，这些方法无法引导模型专注于学习上下文区域中的鲁棒上下文表示，因为它们缺乏明确的机制来感知这些模糊区域，导致这种相似性的次优可判别性。此外，通过模仿临床医生的诊断过程，提供学习区域的视觉证据可以增强检测工具的可解释性，并指导模型生成可靠的决策。在本文中，我们提出了类上下文表征学习与临时视觉解释，以提高肺结节检测。它利用基于查询的机制，通过类上下文向量显式地感知模糊的上下文区域，以指导上下文表示的学习。根据流形假设，该查询将嵌入在特征流形中的上下文信息压缩到一个低维的类上下文潜在空间中，从而过滤冗余以学习判别表示。为了确保这个查询的有效性，我们提出了一个盒引导的实例级方法，使类上下文向量与它们各自类的区别表示保持一致。我们的方法具有可靠的内在可视化效果，增强了模型决策过程的透明度，并为进一步的性能提升提供了学习指导。实验表明，该方法在LUNA16和PN9数据集上的竞争性能指标分别达到了93.38%和65.58%。

{"title":"Box-guided class-contextual representation learning for self-visual lung nodule detection","authors":"Kefu Zhao, Lei Zhang, Le Yi, Xiuyuan Xu","doi":"10.1016/j.bspc.2026.109731","DOIUrl":"10.1016/j.bspc.2026.109731","url":null,"abstract":"<div><div>Detecting lung nodules from computed tomography images has become the routine practice for early lung cancer screening. Conventional methods have made progress in discriminating the high similarity between nodules and non-nodules by extracting contextual information through attention-based or multi-scale mechanisms. However, these methods fail to guide models to focus on learning robust contextual representations in contextual regions as they lack explicit mechanisms to perceive these ambiguous regions, leading to a suboptimal discriminability for such similarity. Moreover, by mimicking clinicians’ diagnostic processes, providing visual evidence of learned regions can enhance the interpretability of detection tools and guide models to generate reliable decisions. In this paper, we propose class-contextual representation learning with ante-hoc visual interpretations to enhance lung nodule detection. It leverages a query-based mechanism via class-contextual vectors to explicitly perceive ambiguous contextual regions for guiding the learning of contextual representations. According to the manifold assumption, this query compresses the contextual information embedded in the feature manifold into a low-dimensional, class-contextual latent space, thereby filtering redundancy for learning discriminative representations. To ensure the efficacy of this query, we propose a box-guided instance-level method that enables class-contextual vectors align with discriminative representations of their respective classes. Our method exhibits a reliable intrinsic visualization effect, which enhances the transparency of the model’s decision-making process and provides learning guidance for further performance gains. Experiments demonstrate that this method achieves competition performance metrics of 93.38% and 65.58% on the LUNA16 and PN9 datasets, respectively.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109731"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174417","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

AMF-MedIT: An efficient align-modulation-fusion framework for medical image–tabular data AMF-MedIT：一个有效的医学图像表格数据对齐-调制-融合框架

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109772

Congjing Yu , Jing Ye , Yang Liu , Xiaodong Zhang , Zhiyong Zhang

Multimodal medical analysis combining image and tabular data has gained increasing attention. However, effective fusion remains challenging due to cross-modal discrepancies in feature dimensions and modality contributions, as well as the noise from high-dimensional tabular inputs. To address these problems, we present AMF-MedIT, an efficient Align-Modulation-Fusion framework for medical image and tabular data integration, particularly under data-scarce conditions. Built upon a self-supervised learning strategy, we introduce the Adaptive Modulation and Fusion (AMF) module, a novel, streamlined fusion paradigm that explicitly regulates the mismatch between modality confidence and representation dominance. Instead of relying on data-hungry learning, the AMF module incorporates modality confidence priors to guide adaptive contribution allocation, while harmonizing unimodal feature dimensionality and magnitude through feature modulation. To enhance tabular representation learning under noisy and limited data conditions, we adopt FT-Mamba as the tabular encoder within the proposed framework, leveraging its selective mechanism to extract discriminative features. Extensive experiments on both clean and clinically noisy datasets demonstrate that AMF-MedIT achieves superior accuracy, robustness, and data efficiency across multimodal classification tasks. Interpretability analyses further reveal how FT-Mamba shapes multimodal pretraining and enhances the image encoder’s attention, highlighting the practical value of our framework for reliable and efficient clinical artificial intelligence applications. The code is available at https://github.com/Jasmine-ycj/AMF-MedIT.git.

结合图像和表格数据的多模态医学分析越来越受到人们的关注。然而，由于特征维度和模态贡献的跨模态差异，以及来自高维表格输入的噪声，有效的融合仍然具有挑战性。为了解决这些问题，我们提出了AMF-MedIT，这是一个有效的对齐-调制-融合框架，用于医学图像和表格数据集成，特别是在数据稀缺的条件下。在自监督学习策略的基础上，我们引入了自适应调制和融合（AMF）模块，这是一种新颖的流线型融合范式，可以明确调节模态置信度和表征优势之间的不匹配。AMF模块不依赖于数据饥渴型学习，而是采用模态置信度先验来指导自适应贡献分配，同时通过特征调制协调单模态特征维度和大小。为了增强在噪声和有限数据条件下的表格表示学习，我们在提出的框架中采用FT-Mamba作为表格编码器，利用其选择机制提取判别特征。在干净和临床噪声数据集上进行的大量实验表明，AMF-MedIT在多模态分类任务中实现了卓越的准确性、鲁棒性和数据效率。可解释性分析进一步揭示了FT-Mamba如何塑造多模态预训练并增强图像编码器的注意力，突出了我们的框架在可靠和高效的临床人工智能应用中的实用价值。代码可在https://github.com/Jasmine-ycj/AMF-MedIT.git上获得。

{"title":"AMF-MedIT: An efficient align-modulation-fusion framework for medical image–tabular data","authors":"Congjing Yu , Jing Ye , Yang Liu , Xiaodong Zhang , Zhiyong Zhang","doi":"10.1016/j.bspc.2026.109772","DOIUrl":"10.1016/j.bspc.2026.109772","url":null,"abstract":"<div><div>Multimodal medical analysis combining image and tabular data has gained increasing attention. However, effective fusion remains challenging due to cross-modal discrepancies in feature dimensions and modality contributions, as well as the noise from high-dimensional tabular inputs. To address these problems, we present AMF-MedIT, an efficient Align-Modulation-Fusion framework for medical image and tabular data integration, particularly under data-scarce conditions. Built upon a self-supervised learning strategy, we introduce the Adaptive Modulation and Fusion (AMF) module, a novel, streamlined fusion paradigm that explicitly regulates the mismatch between modality confidence and representation dominance. Instead of relying on data-hungry learning, the AMF module incorporates modality confidence priors to guide adaptive contribution allocation, while harmonizing unimodal feature dimensionality and magnitude through feature modulation. To enhance tabular representation learning under noisy and limited data conditions, we adopt FT-Mamba as the tabular encoder within the proposed framework, leveraging its selective mechanism to extract discriminative features. Extensive experiments on both clean and clinically noisy datasets demonstrate that AMF-MedIT achieves superior accuracy, robustness, and data efficiency across multimodal classification tasks. Interpretability analyses further reveal how FT-Mamba shapes multimodal pretraining and enhances the image encoder’s attention, highlighting the practical value of our framework for reliable and efficient clinical artificial intelligence applications. The code is available at <span><span>https://github.com/Jasmine-ycj/AMF-MedIT.git</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109772"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174644","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

MultiScaleSegNet: A novel framework for multi-modal brain tumor segmentation MultiScaleSegNet：一种新的多模态脑肿瘤分割框架

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109786

Syed Fakhar Bilal , Jianqiang Li , Jun Qian , Saqib Ali , Muhammad Arif , Baolin Zhu , Lijun Zhao

Accurate segmentation of brain tumors from multi-modal MRI is critical for diagnosis and treatment, but remains challenging due to heterogeneous tumor morphology, ambiguous boundaries, and the need to integrate both local details and global context. To address these challenges, we propose MultiScaleSegNet, a novel encoder–decoder framework that synergistically integrates a Swin Transformer encoder with a DenseNet-based decoder. Our model introduces three key components: (1) a Dual-Path Attention mechanism for feature extraction (DPA-MFE) that preserves spatial details while modeling long-range dependencies; (2) an Attention-based Feature Enhanced Network (AFENet) at the bottleneck to recalibrate features channel-wise and spatially; and (3) a Cross-Feature Refinement (CFR) block that expands the receptive field using dilated convolutions. The decoder further leverages CFR-refined skip connections to recover precise boundary information. Trained with a hybrid BCE-Dice loss on BraTS 2020 and BraTS 2021 datasets, our model achieves state-of-the-art performance, with average Dice scores of 0.972 and 0.987, respectively. Extensive experiments, including ablation studies and comparisons with existing methods, demonstrate that MultiScaleSegNet provides a robust and accurate solution for brain tumor segmentation, offering a strong foundation for clinical applications.

从多模态MRI中准确分割脑肿瘤对诊断和治疗至关重要，但由于肿瘤形态异质性、边界模糊以及需要整合局部细节和全局背景，仍然具有挑战性。为了应对这些挑战，我们提出了MultiScaleSegNet，这是一种新颖的编码器-解码器框架，它协同集成了Swin Transformer编码器和基于densenet的解码器。我们的模型引入了三个关键组件：(1)用于特征提取的双路径注意机制（DPA-MFE），在建模远程依赖关系的同时保留空间细节；(2)在瓶颈处使用基于注意力的特征增强网络（AFENet）对特征进行通道和空间上的重新校准；(3)使用扩张卷积扩展接受野的交叉特征细化（CFR）块。解码器进一步利用cfr精炼的跳过连接来恢复精确的边界信息。在BraTS 2020和BraTS 2021数据集上使用混合BCE-Dice损失进行训练，我们的模型达到了最先进的性能，平均Dice得分分别为0.972和0.987。广泛的实验，包括消融研究和与现有方法的比较，表明MultiScaleSegNet为脑肿瘤分割提供了一个强大而准确的解决方案，为临床应用提供了坚实的基础。

{"title":"MultiScaleSegNet: A novel framework for multi-modal brain tumor segmentation","authors":"Syed Fakhar Bilal , Jianqiang Li , Jun Qian , Saqib Ali , Muhammad Arif , Baolin Zhu , Lijun Zhao","doi":"10.1016/j.bspc.2026.109786","DOIUrl":"10.1016/j.bspc.2026.109786","url":null,"abstract":"<div><div>Accurate segmentation of brain tumors from multi-modal MRI is critical for diagnosis and treatment, but remains challenging due to heterogeneous tumor morphology, ambiguous boundaries, and the need to integrate both local details and global context. To address these challenges, we propose MultiScaleSegNet, a novel encoder–decoder framework that synergistically integrates a Swin Transformer encoder with a DenseNet-based decoder. Our model introduces three key components: (1) a Dual-Path Attention mechanism for feature extraction (DPA-MFE) that preserves spatial details while modeling long-range dependencies; (2) an Attention-based Feature Enhanced Network (AFENet) at the bottleneck to recalibrate features channel-wise and spatially; and (3) a Cross-Feature Refinement (CFR) block that expands the receptive field using dilated convolutions. The decoder further leverages CFR-refined skip connections to recover precise boundary information. Trained with a hybrid BCE-Dice loss on BraTS 2020 and BraTS 2021 datasets, our model achieves state-of-the-art performance, with average Dice scores of 0.972 and 0.987, respectively. Extensive experiments, including ablation studies and comparisons with existing methods, demonstrate that MultiScaleSegNet provides a robust and accurate solution for brain tumor segmentation, offering a strong foundation for clinical applications.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109786"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174415","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

Analysis of optical tweezers single-molecule force spectroscopy based on a signal-enhanced denoising model 基于信号增强去噪模型的光镊单分子力谱分析

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109761

Linyao Chen, Jingru Sun, Le Wang, Hao Huang, Yanghui Li

Optical Tweezers-based Single-Molecule Force Spectroscopy enables nanoscale investigation of biological molecules but is plagued by noise that interferes with Force-Distance Curves (FDCs). This study presents an automated analysis method comprising a Sliding Slice Denoiser (SSD) and an FDC Analysis Module. The SSD employs adaptive segmentation and a neural network integrated with Inception blocks and Self-Attention modules for denoising, then reconstructs high signal‑to‑noise ratios (SNR) FDCs. The module performs folding event quantification, site localization, and Worm-Like Chain fitting to extract biophysical parameters. Tests on single-fold Deoxyribonucleic Acid (DNA) hairpins show improved SNR, with the distance signal increasing from 21.8 dB to 53.6 dB and the force signal from 30.9 dB to 53.2 dB. Mean Absolute Errors of fold site are low, at approximately 0.097 pN for force and 0.73 nm for distance, with Coefficient of Determination exceeding 0.97. For 1 to 6 folds simulated FDCs, the overall fold count prediction accuracy reaches 99%.

基于光镊子的单分子力谱技术使生物分子的纳米级研究成为可能，但它受到干扰力-距离曲线（fdc）的噪声的困扰。本研究提出一种自动分析方法，包括滑动切片去噪（SSD）和FDC分析模块。SSD采用自适应分割和集成了Inception模块和自关注模块的神经网络进行去噪，然后重建高信噪比（SNR） fdc。该模块执行折叠事件量化、位点定位和蠕虫样链拟合以提取生物物理参数。单次DNA发夹测试表明，距离信号从21.8 dB增加到53.6 dB，力信号从30.9 dB增加到53.2 dB，信噪比有所提高。折叠位置的平均绝对误差较低，力的平均绝对误差约为0.097 pN，距离的平均绝对误差约为0.73 nm，决定系数超过0.97。对于1 ~ 6次的模拟fdc，总体的折叠数预测精度达到99%。

{"title":"Analysis of optical tweezers single-molecule force spectroscopy based on a signal-enhanced denoising model","authors":"Linyao Chen, Jingru Sun, Le Wang, Hao Huang, Yanghui Li","doi":"10.1016/j.bspc.2026.109761","DOIUrl":"10.1016/j.bspc.2026.109761","url":null,"abstract":"<div><div>Optical Tweezers-based Single-Molecule Force Spectroscopy enables nanoscale investigation of biological molecules but is plagued by noise that interferes with Force-Distance Curves (FDCs). This study presents an automated analysis method comprising a Sliding Slice Denoiser (SSD) and an FDC Analysis Module. The SSD employs adaptive segmentation and a neural network integrated with Inception blocks and Self-Attention modules for denoising, then reconstructs high signal‑to‑noise ratios (SNR) FDCs. The module performs folding event quantification, site localization, and Worm-Like Chain fitting to extract biophysical parameters. Tests on single-fold Deoxyribonucleic Acid (DNA) hairpins show improved SNR, with the distance signal increasing from 21.8 dB to 53.6 dB and the force signal from 30.9 dB to 53.2 dB. Mean Absolute Errors of fold site are low, at approximately 0.097 pN for force and 0.73 nm for distance, with Coefficient of Determination exceeding 0.97. For 1 to 6 folds simulated FDCs, the overall fold count prediction accuracy reaches 99%.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109761"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174597","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

Analysis of bio-plausible spiking neural networks for motor imagery recognition tasks 生物似是而非的脉冲神经网络用于运动图像识别任务的分析

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109500

Xiuqing Wang , Yunpeng Yang , Qingru Li , Xiaoya Ye , Yang An , Qiuting Li

Brain-computer interface (BCI) is one of the critical aspects of human–computer interaction (HCI). Motor imagery (MI) EEG-based BCI has great application potential for assisting disabled people in motor function reconstruction and restoration. To fulfill the given tasks based on MI-EEG signals successfully, accurately decoding MI-EEG signals is significant. Although deep learning (DL) can be used to analyze EEG signals, it still lacks interpretability. Compared with traditional artificial neural networks (ANNs), spiking neural networks (SNNs) use spiking neurons which are bio-plausible for the neurons in the brain to communicate with each other, and analyze EEG signals by spike trains with better bio-interpretability. Aims at providing an effective model with good stability to analyze EEG-based MI information, we propose a deep spiking convolutional neural network based on a self-attention (DSCNN-SA) mechanism for EEG-based MI recognition. The DSCNN-SA model is first used to extract the high-level spatio-temporal features of the EEG signal and transform the extracted features into spike trains as input to the deep spiking convolutional neural networks (DSCNNs), and analyze the EEG signal for MI according to the firing patterns of spiking neurons. The classification of 2-class motor imagery tasks is evaluated by the BCI Competition IV-2a and IV-2b datasets, and the average accuracy of the DSCNN-SA model is 85.53% and 81.52% respectively, which is better than that of the comparison models, such as KNN, MLP, DCNN-SA, ECCSP-TB2B, CNN-SAE, EEGNet, KLD and STNN, etc. The experimental results validate that the DSCNN-SA model is suitable for EEG-based MI recognition.

脑机接口（BCI）是人机交互（HCI）的一个重要方面。基于脑电信号的脑机接口在辅助残疾人运动功能重建和恢复方面具有很大的应用潜力。为了成功地完成基于MI-EEG信号的给定任务，对MI-EEG信号进行准确解码至关重要。虽然深度学习（DL）可以用于分析脑电图信号，但它仍然缺乏可解释性。与传统的人工神经网络（ann）相比，snn利用脑内神经元之间具有生物合理性的spike神经元相互通信，通过具有更好生物可解释性的spike序列分析脑电信号。为了提供一种稳定有效的模型来分析基于脑电图的心梗信息，我们提出了一种基于自注意机制的深度尖峰卷积神经网络（DSCNN-SA）用于基于脑电图的心梗识别。首先利用DSCNN-SA模型提取脑电信号的高层时空特征，并将提取的特征转化为尖峰序列，作为深度尖峰卷积神经网络（dscnn）的输入，根据尖峰神经元的放电模式对脑电信号进行MI分析。利用BCI Competition IV-2a和IV-2b数据集对2类运动图像任务的分类进行评估，DSCNN-SA模型的平均准确率分别为85.53%和81.52%，优于KNN、MLP、DCNN-SA、ECCSP-TB2B、CNN-SAE、EEGNet、KLD和STNN等比较模型。实验结果验证了DSCNN-SA模型适用于基于脑电图的脑梗死识别。

{"title":"Analysis of bio-plausible spiking neural networks for motor imagery recognition tasks","authors":"Xiuqing Wang , Yunpeng Yang , Qingru Li , Xiaoya Ye , Yang An , Qiuting Li","doi":"10.1016/j.bspc.2026.109500","DOIUrl":"10.1016/j.bspc.2026.109500","url":null,"abstract":"<div><div>Brain-computer interface (BCI) is one of the critical aspects of human–computer interaction (HCI). Motor imagery (MI) EEG-based BCI has great application potential for assisting disabled people in motor function reconstruction and restoration. To fulfill the given tasks based on MI-EEG signals successfully, accurately decoding MI-EEG signals is significant. Although deep learning (DL) can be used to analyze EEG signals, it still lacks interpretability. Compared with traditional artificial neural networks (ANNs), spiking neural networks (SNNs) use spiking neurons which are bio-plausible for the neurons in the brain to communicate with each other, and analyze EEG signals by spike trains with better bio-interpretability. Aims at providing an effective model with good stability to analyze EEG-based MI information, we propose a deep spiking convolutional neural network based on a self-attention (DSCNN-SA) mechanism for EEG-based MI recognition. The DSCNN-SA model is first used to extract the high-level spatio-temporal features of the EEG signal and transform the extracted features into spike trains as input to the deep spiking convolutional neural networks (DSCNNs), and analyze the EEG signal for MI according to the firing patterns of spiking neurons. The classification of 2-class motor imagery tasks is evaluated by the BCI Competition IV-2a and IV-2b datasets, and the average accuracy of the DSCNN-SA model is 85.53% and 81.52% respectively, which is better than that of the comparison models, such as KNN, MLP, DCNN-SA, ECCSP-TB2B, CNN-SAE, EEGNet, KLD and STNN, etc. The experimental results validate that the DSCNN-SA model is suitable for EEG-based MI recognition.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109500"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174640","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

MLFormer: Mamba-like linear attention with hierarchical context fusion for efficient medical image segmentation MLFormer：基于分层上下文融合的类曼巴线性注意力的高效医学图像分割

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109745

Xueping Liu , Zelong Yu , Youru Li , Shi Bai , Na Li , Yu Bai , Silu Ding

Recent advances in medical imaging produce high-resolution data but also exacerbate key challenges, including anatomical variability, ambiguous boundaries, low contrast, and signal degradation. Traditional segmentation methods struggle to address these challenges, making deep learning approaches, particularly U-shaped architectures, increasingly prominent. CNNs are limited by local receptive fields, whereas ViTs incur quadratic complexity on large images. To gap these issues, we present MLFormer, a U-shaped network that couples a Mamba-like linear-attention block (MLLA) with a transformer-based context bridge to efficiently fuse global context and local structure. At its core, MLLA combines a kernelized linear-attention formulation with Mamba-inspired selective gating to retain full token-level parallelism while overcoming the expressivity limits of conventional linear attention. We further redesign the transformer context bridge to strengthen the fusion of local details and global semantics within skip connections. We conduct extensive evaluations on five datasets which cover three imaging modalities and diverse anatomies. Experiments show that our proposed method achieves accurate and data-efficient segmentation while maintaining favorable computational efficiency. Our code is released at https://github.com/MEAI-SAU/MLFormer.

医学成像的最新进展产生了高分辨率数据，但也加剧了关键挑战，包括解剖变异性、模糊边界、低对比度和信号退化。传统的分割方法难以应对这些挑战，这使得深度学习方法，特别是u型架构日益突出。cnn受局部接受域的限制，而vit在大图像上产生二次复杂度。为了解决这些问题，我们提出了MLFormer，这是一个u形网络，它将一个类似曼巴的线性注意块（MLLA）与一个基于转换器的上下文桥结合在一起，以有效地融合全局上下文和局部结构。在其核心，MLLA结合了核线性注意公式与曼巴启发的选择性门控，以保留完整的令牌级并行性，同时克服了传统线性注意的表达能力限制。我们进一步重新设计了转换器上下文桥，以加强跳过连接中局部细节和全局语义的融合。我们对涵盖三种成像方式和不同解剖结构的五个数据集进行了广泛的评估。实验表明，该方法在保持良好的计算效率的同时，实现了准确、高效的数据分割。我们的代码发布在https://github.com/MEAI-SAU/MLFormer。

{"title":"MLFormer: Mamba-like linear attention with hierarchical context fusion for efficient medical image segmentation","authors":"Xueping Liu , Zelong Yu , Youru Li , Shi Bai , Na Li , Yu Bai , Silu Ding","doi":"10.1016/j.bspc.2026.109745","DOIUrl":"10.1016/j.bspc.2026.109745","url":null,"abstract":"<div><div>Recent advances in medical imaging produce high-resolution data but also exacerbate key challenges, including anatomical variability, ambiguous boundaries, low contrast, and signal degradation. Traditional segmentation methods struggle to address these challenges, making deep learning approaches, particularly U-shaped architectures, increasingly prominent. CNNs are limited by local receptive fields, whereas ViTs incur quadratic complexity on large images. To gap these issues, we present MLFormer, a U-shaped network that couples a Mamba-like linear-attention block (MLLA) with a transformer-based context bridge to efficiently fuse global context and local structure. At its core, MLLA combines a kernelized linear-attention formulation with Mamba-inspired selective gating to retain full token-level parallelism while overcoming the expressivity limits of conventional linear attention. We further redesign the transformer context bridge to strengthen the fusion of local details and global semantics within skip connections. We conduct extensive evaluations on five datasets which cover three imaging modalities and diverse anatomies. Experiments show that our proposed method achieves accurate and data-efficient segmentation while maintaining favorable computational efficiency. Our code is released at <span><span>https://github.com/MEAI-SAU/MLFormer</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109745"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174645","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

LAVAE-Net: A deep LSTM-attention VAE framework for detecting ECG anomalies LAVAE-Net：用于心电异常检测的深度lstm -注意VAE框架

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109755

Hao Shi , Zhuoying Jiang

Electrocardiogram (ECG) abnormality detection is crucial for the early diagnosis of cardiovascular diseases. Traditional supervised-learning methods, however, requires abundant labeled abnormal samples, which are clinically scarce and costly to annotate. To overcome this limitation, we propose LAVAE-Net—an unsupervised ECG anomaly-detection model that fuses a Variational Autoencoder (VAE) with a Long Short-Term Memory (LSTM) network enhanced by an attention mechanism. LAVAE-Net trains exclusively on normal ECG data and identifies anomalies by learning the intrinsic latent distribution of normal cardiac dynamics. Architecturally, the encoder employs an LSTM backbone to capture long-term temporal dependencies, while an attention module adaptively highlights diagnostically salient features. The VAE encoder then maps the sequential ECG signal into a low-dimensional latent space and reparameterizes it into Gaussian-distributed latent variables. A GRU-based decoder reconstructs the original sequence from these latent variables, enabling the model to quantify deviations from the learned normal manifold. Training is driven by a composite loss that balances reconstruction error with Kullback–Leibler divergence, ensuring that latent representations remain both faithful and regularized. Evaluated on the public ECG5000 dataset, LAVAE-Net attains an F1 score of 99.4% for anomaly detection and maintains high precision for both normal and abnormal instances. Consequently, by eliminating the need for labeled anomalies, the model markedly improves clinical practicality and generalization thus providing a robust foundation for intelligent real-time ECG monitoring systems.

心电图异常检测对心血管疾病的早期诊断至关重要。然而，传统的监督学习方法需要大量标记的异常样本，这些样本在临床上是稀缺的，并且注释成本很高。为了克服这一限制，我们提出了lavae - net -一种无监督ECG异常检测模型，该模型融合了变分自编码器（VAE）和由注意机制增强的长短期记忆（LSTM）网络。LAVAE-Net只训练正常的心电数据，并通过学习正常心脏动力学的内在潜在分布来识别异常。在架构上，编码器采用LSTM主干来捕获长期时间依赖性，而注意力模块自适应地突出诊断显着特征。然后，VAE编码器将序列心电信号映射到低维潜在空间，并将其重新参数化为高斯分布的潜在变量。基于gru的解码器从这些潜在变量重建原始序列，使模型能够量化与学习到的正态流形的偏差。训练由复合损失驱动，该损失平衡了重建误差和Kullback-Leibler散度，确保潜在表示保持忠实和正则化。在公开的ECG5000数据集上进行评估，LAVAE-Net异常检测的F1得分达到99.4%，对正常和异常实例都保持了较高的精度。因此，通过消除对标记异常的需要，该模型显着提高了临床实用性和通用性，从而为智能实时心电监测系统提供了坚实的基础。

{"title":"LAVAE-Net: A deep LSTM-attention VAE framework for detecting ECG anomalies","authors":"Hao Shi , Zhuoying Jiang","doi":"10.1016/j.bspc.2026.109755","DOIUrl":"10.1016/j.bspc.2026.109755","url":null,"abstract":"<div><div>Electrocardiogram (ECG) abnormality detection is crucial for the early diagnosis of cardiovascular diseases. Traditional supervised-learning methods, however, requires abundant labeled abnormal samples, which are clinically scarce and costly to annotate. To overcome this limitation, we propose LAVAE-Net—an unsupervised ECG anomaly-detection model that fuses a Variational Autoencoder (VAE) with a Long Short-Term Memory (LSTM) network enhanced by an attention mechanism. LAVAE-Net trains exclusively on normal ECG data and identifies anomalies by learning the intrinsic latent distribution of normal cardiac dynamics. Architecturally, the encoder employs an LSTM backbone to capture long-term temporal dependencies, while an attention module adaptively highlights diagnostically salient features. The VAE encoder then maps the sequential ECG signal into a low-dimensional latent space and reparameterizes it into Gaussian-distributed latent variables. A GRU-based decoder reconstructs the original sequence from these latent variables, enabling the model to quantify deviations from the learned normal manifold. Training is driven by a composite loss that balances reconstruction error with Kullback–Leibler divergence, ensuring that latent representations remain both faithful and regularized. Evaluated on the public ECG5000 dataset, LAVAE-Net attains an F1 score of 99.4% for anomaly detection and maintains high precision for both normal and abnormal instances. Consequently, by eliminating the need for labeled anomalies, the model markedly improves clinical practicality and generalization thus providing a robust foundation for intelligent real-time ECG monitoring systems.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109755"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174483","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

MABNSleepNet: A multi-view self-attention architecture based on brain network for neonatal sleep staging MABNSleepNet：一种基于大脑网络的新生儿睡眠分期多视图自注意架构

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-02-04 DOI: 10.1016/j.bspc.2026.109651

Ligang Zhou , Yan Xu , Yalin Wang , Xiaoyu Chen , Laishuan Wang , Wei Chen , Chen Chen

Existing deep learning-based automatic sleep staging methods for neonates predominantly rely on CNN or RNN architectures, often overlooking information on inter-channel spatial and inter-stage temporal dynamics. Although GNN-based and graph-based methods can leverage such information, they solely focus on functional connectivity from linear relationships instead of incorporating effective connectivity from information flow. Thus, a multi-view self-attention based on the brain network method is proposed for neonatal sleep staging (MABNSleepNet), which consists of a Brain Network Extractor (BNE), a Context-attaching Module (CAM), a Multi-view Self-attention Module (MSAM), and a Feature-aggregating Module (FAM). The proposed method leverages multi-view self-attention and brain connectivity analysis to comprehensively capture both spatial, temporal, and frequency-domain features, while incorporating inter-stage temporal context for accurate neonatal sleep staging. It was validated on a clinical neonatal sleep dataset from the Children’s Hospital of Fudan University (CHFU) with 64 eight-channel EEG recordings. Employing the subject-wise 10-fold validation method with six testing set and 58 training set, the proposed method achieved impressive results on two-stage (wakefulness and sleep) and three-stage (wakefulness, active sleep, and quiet sleep) tasks respectively: with accuracies of 0.920 and 0.862, F1-scores of 0.909 and 0.861, as well as Cohen’s Kappa coefficients of 0.817 and 0.793. Experiment results demonstrate the proposed method’s superiority over most current state-of-the-art methods in neonatal sleep staging, providing a novel avenue for in-depth analysis of sleep staging from the brain connectivity perspective.

现有的基于深度学习的新生儿自动睡眠分期方法主要依赖于CNN或RNN架构，往往忽略了通道间空间和阶段间时间动态的信息。尽管基于gnn和基于图的方法可以利用这些信息，但它们只关注线性关系的功能连通性，而不是信息流的有效连通性。因此，提出了一种基于脑网络的新生儿睡眠分期多视图自注意方法（MABNSleepNet），该方法由脑网络提取器（BNE）、上下文附加模块（CAM）、多视图自注意模块（MSAM）和特征聚合模块（FAM）组成。该方法利用多视角自我注意和大脑连接分析来全面捕捉空间、时间和频域特征，同时结合阶段间时间背景来准确地进行新生儿睡眠分期。在复旦大学儿童医院（CHFU）的64个8通道脑电图记录的临床新生儿睡眠数据集上进行了验证。采用6个测试集和58个训练集的被试10倍验证方法，该方法在两阶段（清醒和睡眠）和三阶段（清醒、活跃和安静睡眠）任务上分别取得了令人印象深刻的结果：准确率分别为0.920和0.862，f1得分分别为0.909和0.861,Cohen’s Kappa系数分别为0.817和0.793。实验结果表明，该方法在新生儿睡眠分期方面优于目前大多数最先进的方法，为从大脑连接角度深入分析睡眠分期提供了新的途径。

{"title":"MABNSleepNet: A multi-view self-attention architecture based on brain network for neonatal sleep staging","authors":"Ligang Zhou , Yan Xu , Yalin Wang , Xiaoyu Chen , Laishuan Wang , Wei Chen , Chen Chen","doi":"10.1016/j.bspc.2026.109651","DOIUrl":"10.1016/j.bspc.2026.109651","url":null,"abstract":"<div><div>Existing deep learning-based automatic sleep staging methods for neonates predominantly rely on CNN or RNN architectures, often overlooking information on inter-channel spatial and inter-stage temporal dynamics. Although GNN-based and graph-based methods can leverage such information, they solely focus on functional connectivity from linear relationships instead of incorporating effective connectivity from information flow. Thus, a multi-view self-attention based on the brain network method is proposed for neonatal sleep staging (MABNSleepNet), which consists of a Brain Network Extractor (BNE), a Context-attaching Module (CAM), a Multi-view Self-attention Module (MSAM), and a Feature-aggregating Module (FAM). The proposed method leverages multi-view self-attention and brain connectivity analysis to comprehensively capture both spatial, temporal, and frequency-domain features, while incorporating inter-stage temporal context for accurate neonatal sleep staging. It was validated on a clinical neonatal sleep dataset from the Children’s Hospital of Fudan University (CHFU) with 64 eight-channel EEG recordings. Employing the subject-wise 10-fold validation method with six testing set and 58 training set, the proposed method achieved impressive results on two-stage (wakefulness and sleep) and three-stage (wakefulness, active sleep, and quiet sleep) tasks respectively: with accuracies of 0.920 and 0.862, F1-scores of 0.909 and 0.861, as well as Cohen’s Kappa coefficients of 0.817 and 0.793. Experiment results demonstrate the proposed method’s superiority over most current state-of-the-art methods in neonatal sleep staging, providing a novel avenue for in-depth analysis of sleep staging from the brain connectivity perspective.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"118 ","pages":"Article 109651"},"PeriodicalIF":4.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174641","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