Biomedical Signal Processing and Control最新文献_第2页

Federated learning for brain tumour detection and classification using improved ShuffleNet-PCNN architecture 基于改进ShuffleNet-PCNN架构的脑肿瘤检测和分类的联邦学习

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-07 DOI: 10.1016/j.bspc.2026.109719

A.Ramesh Khanna, P. Thilakavathy

Brain tumors are a major worldwide health concern, which emphasizes the significance of prompt and accurate diagnosis for efficient treatment planning and management. Differentiating tumors from normal tissues is essential when assessing medical imaging. However, there are a number of challenges with traditional medical imaging techniques for brain tumor detection. Limited datasets, regulations on privacy limiting data sharing, and the requirement for specialized knowledge to correctly analyze medical images could all be obstacles to current approaches. A novel approach called Federated Learning with SNet-PC for Brain Tumor Detection and Classification (FL-SNet-PC) is proposed. This approach utilizes Federated Learning, which incorporates LP-pooled layer-assisted ShuffleNet (LP-SNet) and Parallel Convolutional Neural Network (PCNN) models. During local training, the SNet-PC method is used, which combines the LP-SNet and PCNN architectures. The local training pipeline has several stages, such as preprocessing, segmentation, and feature extraction. Initially, the input image undergoes preprocessing using the Wiener filtering technique to normalize the image. Then, precise segmentation is achieved using the Yeo-Johnson-based Balanced Iterative Reducing and Clustering Using Hierarchies (YJ-BIRCH) algorithm. After segmentation, feature extraction is done, where shape features, Grey-Level Co-Occurrence Matrix (GLCM) features, and Sobel Gradient-based Pyramid Histogram of Gradient Orientation (SG-PHOG) are captured from the segmented image. Once the local training process is completed, they are then sent to a central server for global aggregation. Finally, the global training process aids in detecting and classifying brain tumors effectively.

脑肿瘤是一个主要的全球健康问题，这强调了及时和准确的诊断对有效的治疗计划和管理的重要性。在评估医学影像时，将肿瘤与正常组织区分开来至关重要。然而，传统的医学成像技术在脑肿瘤检测方面存在许多挑战。有限的数据集、限制数据共享的隐私法规，以及对正确分析医学图像的专业知识的要求，都可能成为当前方法的障碍。提出了一种基于SNet-PC的联邦学习脑肿瘤检测与分类方法。这种方法利用了联邦学习，结合了lp池层辅助ShuffleNet （LP-SNet）和并行卷积神经网络（PCNN）模型。在局部训练中，采用了结合LP-SNet和PCNN架构的SNet-PC方法。局部训练管道包括预处理、分割和特征提取等几个阶段。首先，使用维纳滤波技术对输入图像进行预处理，使图像归一化。然后，使用基于yeo - johnson的平衡迭代约简和分层聚类（YJ-BIRCH）算法实现精确分割。分割后进行特征提取，从分割后的图像中获取形状特征、灰度共生矩阵（GLCM）特征和基于Sobel梯度的梯度方向金字塔直方图（SG-PHOG）。一旦本地训练过程完成，它们就会被发送到中央服务器进行全局聚合。最后，全局训练过程有助于有效地检测和分类脑肿瘤。

{"title":"Federated learning for brain tumour detection and classification using improved ShuffleNet-PCNN architecture","authors":"A.Ramesh Khanna, P. Thilakavathy","doi":"10.1016/j.bspc.2026.109719","DOIUrl":"10.1016/j.bspc.2026.109719","url":null,"abstract":"<div><div>Brain tumors are a major worldwide health concern, which emphasizes the significance of prompt and accurate diagnosis for efficient treatment planning and management. Differentiating tumors from normal tissues is essential when assessing medical imaging. However, there are a number of challenges with traditional medical imaging techniques for brain tumor detection. Limited datasets, regulations on privacy limiting data sharing, and the requirement for specialized knowledge to correctly analyze medical images could all be obstacles to current approaches. A novel approach called Federated Learning with SNet-PC for Brain Tumor Detection and Classification (FL-SNet-PC) is proposed. This approach utilizes Federated Learning, which incorporates LP-pooled layer-assisted ShuffleNet (LP-SNet) and Parallel Convolutional Neural Network (PCNN) models. During local training, the SNet-PC method is used, which combines the LP-SNet and PCNN architectures. The local training pipeline has several stages, such as preprocessing, segmentation, and feature extraction. Initially, the input image undergoes preprocessing using the Wiener filtering technique to normalize the image. Then, precise segmentation is achieved using the Yeo-Johnson-based Balanced Iterative Reducing and Clustering Using Hierarchies (YJ-BIRCH) algorithm. After segmentation, feature extraction is done, where shape features, Grey-Level Co-Occurrence Matrix (GLCM) features, and Sobel Gradient-based Pyramid Histogram of Gradient Orientation (SG-PHOG) are captured from the segmented image. Once the local training process is completed, they are then sent to a central server for global aggregation. Finally, the global training process aids in detecting and classifying brain tumors effectively.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109719"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146193102","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

DFF-GAN: Dual-sequence frequency fusion GAN with high frequency enhancement for CT synthesis DFF-GAN：双序列频率融合GAN，用于CT合成的高频增强

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-13 DOI: 10.1016/j.bspc.2026.109844

Gang Yu , Dekang Zhang , Kexin Wei , Shiteng Zhu , Yungang Wang

In clinical practice, multi-sequence MR images provide complementary information that enhances the accuracy of synthetic CT (sCT) generation. However, existing synthesis methods still face the following challenges: first, the fusion strategies for multi-sequence information are limited, failing to fully exploit the complementary nature across modalities; second, deep learning models tend to focus on low-frequency components, resulting in overly smooth sCT that lack critical anatomical details. To address these issues, we propose a Dual-sequence Frequency Fusion Generative Adversarial Network (DFF-GAN), which employs a multi-sequence, multi-frequency fusion strategy and explicitly encourages the model to focus on high-frequency content. The sCT is jointly evaluated by a conventional image discriminator and a high-frequency discriminator, ensuring perceptual realism and rich high-frequency details. Experimental results on pelvic datasets demonstrate that DFF-GAN achieves superior performance compared to state-of-the-art methods in terms of MAE, SSIM, and PSNR, and maintains robustness in the presence of noise or low-quality MR input. Ablation studies confirm the effectiveness of each proposed component, indicating the potential of DFF-GAN in clinical MR-to-CT synthesis applications.

在临床实践中，多序列MR图像提供了补充信息，提高了合成CT （sCT）生成的准确性。然而，现有的合成方法仍然面临着以下挑战：一是多序列信息的融合策略有限，未能充分利用模式间的互补性；其次，深度学习模型倾向于关注低频成分，导致sCT过于平滑，缺乏关键的解剖细节。为了解决这些问题，我们提出了一种双序列频率融合生成对抗网络（DFF-GAN），它采用多序列、多频率融合策略，并明确鼓励模型专注于高频内容。sCT由传统图像鉴别器和高频鉴别器联合评估，保证了感知真实感和丰富的高频细节。盆腔数据集的实验结果表明，与最先进的方法相比，DFF-GAN在MAE、SSIM和PSNR方面取得了更好的性能，并且在存在噪声或低质量MR输入的情况下保持了鲁棒性。消融研究证实了每种提议成分的有效性，表明DFF-GAN在临床mr - ct合成应用中的潜力。

{"title":"DFF-GAN: Dual-sequence frequency fusion GAN with high frequency enhancement for CT synthesis","authors":"Gang Yu , Dekang Zhang , Kexin Wei , Shiteng Zhu , Yungang Wang","doi":"10.1016/j.bspc.2026.109844","DOIUrl":"10.1016/j.bspc.2026.109844","url":null,"abstract":"<div><div>In clinical practice, multi-sequence MR images provide complementary information that enhances the accuracy of synthetic CT (sCT) generation. However, existing synthesis methods still face the following challenges: first, the fusion strategies for multi-sequence information are limited, failing to fully exploit the complementary nature across modalities; second, deep learning models tend to focus on low-frequency components, resulting in overly smooth sCT that lack critical anatomical details. To address these issues, we propose a Dual-sequence Frequency Fusion Generative Adversarial Network (DFF-GAN), which employs a multi-sequence, multi-frequency fusion strategy and explicitly encourages the model to focus on high-frequency content. The sCT is jointly evaluated by a conventional image discriminator and a high-frequency discriminator, ensuring perceptual realism and rich high-frequency details. Experimental results on pelvic datasets demonstrate that DFF-GAN achieves superior performance compared to state-of-the-art methods in terms of MAE, SSIM, and PSNR, and maintains robustness in the presence of noise or low-quality MR input. Ablation studies confirm the effectiveness of each proposed component, indicating the potential of DFF-GAN in clinical MR-to-CT synthesis applications.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109844"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192611","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multimodal deep learning with attention mechanisms for automated detection of lower extremity deep vein thrombosis: Integrating ultrasound, CT, and MRI 多模态深度学习与下肢深静脉血栓自动检测的注意机制：整合超声、CT和MRI

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-10 DOI: 10.1016/j.bspc.2026.109749

Haojie Wang, Yuanhu Jing, Hongyuan Li, Yan Zhang, Congcong Chang, Gongning Shi

This study developed a two-stage cascaded segmentation framework with integrated attention mechanisms (SE/CBAM) that accommodates multimodal inputs from CT, MRI, and ultrasound and demonstrates robustness to missing modalities. The proposed approach was systematically evaluated on data from 322 with deep vein thrombosis (DVT) and compared against representative methods, including nnU-Net, Swin-UNet, UNETR, and MedSAM/EMedSAM. Under standardized preprocessing and five-fold cross-validation, our model achieved a Dice coefficient of 0.873 ± 0.018, IoU of 0.783 ± 0.021, PR-AUC of 0.921, and ROC-AUC of 0.942 on an independent test set, significantly outperforming nnU-Net, Swin-UNet, UNETR, and MedSAM/EMedSAM baselines. Grad-CAM heatmaps showed strong spatial concordance with expert annotations at key anatomical regions. Sensitivity analyses confirmed robustness to hyperparameter variation, noise, and single-modality dropout (<1% fluctuation). To enhance biological interpretability, we integrated transcriptomic (GEO, n = 40), proteomic (PRIDE, n = 30), and metabolomic (MetaboLights, n = 20) datasets, identifying key molecules such as RPS3A, RPL31, and TP53, as well as differential metabolites including oxidized glutathione, succinate, and betaine. These molecular alterations were predominantly enriched in glutathione metabolism, the tricarboxylic acid (TCA) cycle, and inflammation-related pathways (e.g., the IL-17 and TNF signaling axes). Notably, these pathways are directionally consistent with the inflammatory activation and oxidative stress features reflected in the high-risk thrombotic anatomical regions emphasized by the imaging model, thereby providing supportive mechanistic interpretation for the imaging-based recognition results. Collectively, this study establishes a methodological foundation for interpretable AI-driven diagnosis and precision intervention in DVT and demonstrates potential for extension to other vascular diseases.

本研究开发了一个具有综合注意机制（SE/CBAM）的两阶段级联分割框架，该框架可容纳来自CT、MRI和超声的多模态输入，并证明了对缺失模态的鲁棒性。我们对322例深静脉血栓患者的数据进行了系统评估，并与具有代表性的方法（包括nnU-Net、swwin - unet、UNETR和MedSAM/EMedSAM）进行了比较。经过标准化预处理和五重交叉验证，我们的模型在独立测试集上的Dice系数为0.873±0.018，IoU为0.783±0.021，PR-AUC为0.921，ROC-AUC为0.942，显著优于nnU-Net、swun - unet、UNETR和MedSAM/EMedSAM基线。在关键解剖区域，Grad-CAM热图与专家标注的空间一致性较强。敏感性分析证实了对超参数变化、噪声和单模态丢失（<；1%波动）的稳健性。为了提高生物学可解释性，我们整合了转录组学（GEO, n = 40）、蛋白质组学（PRIDE, n = 30）和代谢组学（metabolomics, n = 20）数据集，确定了关键分子，如RPS3A、RPL31和TP53，以及差异代谢物，包括氧化谷胱甘肽、琥珀酸盐和甜菜碱。这些分子改变主要富集于谷胱甘肽代谢、三羧酸（TCA）循环和炎症相关途径（例如，IL-17和TNF信号轴）。值得注意的是，这些途径与成像模型强调的高危血栓解剖区域所反映的炎症激活和氧化应激特征在方向上是一致的，从而为基于成像的识别结果提供了支持性的机制解释。总的来说，本研究为可解释的人工智能驱动的DVT诊断和精确干预奠定了方法学基础，并展示了扩展到其他血管疾病的潜力。

{"title":"Multimodal deep learning with attention mechanisms for automated detection of lower extremity deep vein thrombosis: Integrating ultrasound, CT, and MRI","authors":"Haojie Wang, Yuanhu Jing, Hongyuan Li, Yan Zhang, Congcong Chang, Gongning Shi","doi":"10.1016/j.bspc.2026.109749","DOIUrl":"10.1016/j.bspc.2026.109749","url":null,"abstract":"<div><div>This study developed a two-stage cascaded segmentation framework with integrated attention mechanisms (SE/CBAM) that accommodates multimodal inputs from CT, MRI, and ultrasound and demonstrates robustness to missing modalities. The proposed approach was systematically evaluated on data from 322 with deep vein thrombosis (DVT) and compared against representative methods, including nnU-Net, Swin-UNet, UNETR, and MedSAM/EMedSAM. Under standardized preprocessing and five-fold cross-validation, our model achieved a Dice coefficient of 0.873 ± 0.018, IoU of 0.783 ± 0.021, PR-AUC of 0.921, and ROC-AUC of 0.942 on an independent test set, significantly outperforming nnU-Net, Swin-UNet, UNETR, and MedSAM/EMedSAM baselines. Grad-CAM heatmaps showed strong spatial concordance with expert annotations at key anatomical regions. Sensitivity analyses confirmed robustness to hyperparameter variation, noise, and single-modality dropout (<1% fluctuation). To enhance biological interpretability, we integrated transcriptomic (GEO, n = 40), proteomic (PRIDE, n = 30), and metabolomic (MetaboLights, n = 20) datasets, identifying key molecules such as RPS3A, RPL31, and TP53, as well as differential metabolites including oxidized glutathione, succinate, and betaine. These molecular alterations were predominantly enriched in glutathione metabolism, the tricarboxylic acid (TCA) cycle, and inflammation-related pathways (e.g., the IL-17 and TNF signaling axes). Notably, these pathways are directionally consistent with the inflammatory activation and oxidative stress features reflected in the high-risk thrombotic anatomical regions emphasized by the imaging model, thereby providing supportive mechanistic interpretation for the imaging-based recognition results. Collectively, this study establishes a methodological foundation for interpretable AI-driven diagnosis and precision intervention in DVT and demonstrates potential for extension to other vascular diseases.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109749"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192858","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

TFR-GANNet: Multi-channel time–frequency ridge fusion and CWGAN-GP for sleep arousal identification TFR-GANNet：多通道时频脊融合和CWGAN-GP用于睡眠觉醒识别

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-10 DOI: 10.1016/j.bspc.2026.109846

Wenjuan Lu , Zhongxing Liu , Yu Chen , Jinxi Wang

The precise identification of sleep arousal events is critical for sleep quality assessment and diagnosis of sleep disorders. Current methods face three major limitations: (1) the lack of an effective multi-channel signal fusion method, (2) inadequate handling of class-imbalanced datasets, and (3) reliance on expert annotations, which hinder the clinical applicability of existing methods. To address these limitations, an integrated method named TFR-GANNet was proposed, which incorporates three modules for multi-channel data fusion, data augmentation, and sleep arousal identification. First, the data fusion module implements a novel time–frequency ridge-based multi-channel signal fusion (TFR-MSF) strategy to construct a global time–frequency ridge index map (GTFRIM) by extracting and integrating channel-specific ridge features, enabling physiologically meaningful signal representation. Subsequently, the data augmentation module utilizes a conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP) to synthesize GTFRIM samples, which effectively balances the data distribution across different classes. Finally, the sleep arousal identification module consists of a dense convolutional network integrated with a multi-head attention mechanism (DMNet), enabling accurate identification of sleep arousal events from GTFRIMs. On the PhysioNet 2018 sleep dataset, TFR-GANNet achieved state-of-the-art performance compared with existing methods, with an accuracy of 0.9556, an AUROC of 0.9699, an F1-score of 0.9575, and an AUPRC of 0.8751. Extensive ablation studies confirmed the individual contributions of each module. The proposed framework advances sleep analysis through an integrated solution to key challenges in data fusion, class imbalance, and annotation scarcity, paving the way for robust applications in clinical and research domains.

准确识别睡眠唤醒事件对睡眠质量评估和睡眠障碍的诊断至关重要。目前的方法面临三个主要的局限性：(1)缺乏有效的多通道信号融合方法；(2)对类别不平衡数据集的处理不足；(3)依赖专家注释，阻碍了现有方法的临床适用性。为了解决这些问题，提出了一种TFR-GANNet集成方法，该方法包括多通道数据融合、数据增强和睡眠唤醒识别三个模块。首先，数据融合模块实现了一种新颖的基于时频脊的多通道信号融合（TFR-MSF）策略，通过提取和整合通道特定的脊特征来构建全局时频脊指数图（GTFRIM），从而实现有生理意义的信号表示。随后，数据增强模块利用带梯度惩罚的条件Wasserstein生成对抗网络（CWGAN-GP）合成GTFRIM样本，有效平衡了不同类别之间的数据分布。最后，睡眠唤醒识别模块由一个与多头注意机制（DMNet）集成的密集卷积网络组成，能够准确识别来自gtfram的睡眠唤醒事件。在PhysioNet 2018睡眠数据集上，与现有方法相比，TFR-GANNet取得了最先进的性能，准确率为0.9556，AUROC为0.9699，f1得分为0.9575，AUPRC为0.8751。广泛的消融研究证实了每个模块的个人贡献。提出的框架通过集成解决数据融合、类不平衡和注释稀缺性等关键挑战来推进睡眠分析，为临床和研究领域的强大应用铺平了道路。

{"title":"TFR-GANNet: Multi-channel time–frequency ridge fusion and CWGAN-GP for sleep arousal identification","authors":"Wenjuan Lu , Zhongxing Liu , Yu Chen , Jinxi Wang","doi":"10.1016/j.bspc.2026.109846","DOIUrl":"10.1016/j.bspc.2026.109846","url":null,"abstract":"<div><div>The precise identification of sleep arousal events is critical for sleep quality assessment and diagnosis of sleep disorders. Current methods face three major limitations: (1) the lack of an effective multi-channel signal fusion method, (2) inadequate handling of class-imbalanced datasets, and (3) reliance on expert annotations, which hinder the clinical applicability of existing methods. To address these limitations, an integrated method named TFR-GANNet was proposed, which incorporates three modules for multi-channel data fusion, data augmentation, and sleep arousal identification. First, the data fusion module implements a novel time–frequency ridge-based multi-channel signal fusion (TFR-MSF) strategy to construct a global time–frequency ridge index map (GTFRIM) by extracting and integrating channel-specific ridge features, enabling physiologically meaningful signal representation. Subsequently, the data augmentation module utilizes a conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP) to synthesize GTFRIM samples, which effectively balances the data distribution across different classes. Finally, the sleep arousal identification module consists of a dense convolutional network integrated with a multi-head attention mechanism (DMNet), enabling accurate identification of sleep arousal events from GTFRIMs. On the PhysioNet 2018 sleep dataset, TFR-GANNet achieved state-of-the-art performance compared with existing methods, with an accuracy of 0.9556, an AUROC of 0.9699, an F1-score of 0.9575, and an AUPRC of 0.8751. Extensive ablation studies confirmed the individual contributions of each module. The proposed framework advances sleep analysis through an integrated solution to key challenges in data fusion, class imbalance, and annotation scarcity, paving the way for robust applications in clinical and research domains.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109846"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192855","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

Parkinson’s disease detection from resting-state EEG using a Transformer model with Multi-Head Attention explanations 基于多头注意解释的变压器模型的静息状态脑电图帕金森病检测

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-09 DOI: 10.1016/j.bspc.2026.109811

Valeriana Mancazzo , Elena Sibilano , Domenico Buongiorno, Raffaele Carli, Vitoantonio Bevilacqua, Antonio Brunetti

Despite the proven potential of using Deep Learning (DL) models based on electroencephalographic (EEG) signals to detect neurological disorders like Parkinson’s Disease (PD), their adoption in clinical practice is limited due to insufficient reliability and generalizability. We propose an interpretable end-to-end DL framework leveraging the Multi-Head Attention (MHA) component of Transformers to classify EEG signals of 100 PD patients and 79 control subjects across three public resting-state EEG datasets. A systematic interpretability approach, including embedding visualization and MHA-based temporal and spectral analysis through statistical tests, is proposed to enhance the identification of discriminative biomarkers. Experimental findings across a large, multi-centric cohort of subjects demonstrated the framework’s capability to detect meaningful EEG patterns in the frequency intervals of interest.

Interpretability analysis revealed that MHA focused on specific temporal patches in the input signal, which correlated to the classification outcomes. Spectral analysis identified significant power differences in Theta and Beta bands, capturing neural patterns of cognitive and motor dysfunction in PD. Furthermore, attention-guided segmentation improved the sensitivity of spectral biomarkers, such as Alpha/Theta ratio and Beta relative power, consistent with prior literature. Moreover, the proposed approach yielded the highest epoch-level mean AUCs of

0.91 \pm 0.01

on Theta,

0.84 \pm 0.03

on Alpha, and

0.81 \pm 0.01

on All-band, achieving state-of-the-art performances while also demonstrating robustness to heterogeneous data.

尽管使用基于脑电图（EEG）信号的深度学习（DL）模型来检测帕金森病（PD）等神经系统疾病已被证明具有潜力，但由于可靠性和通用性不足，它们在临床实践中的应用受到限制。我们提出了一个可解释的端到端深度学习框架，利用变压器的多头注意（MHA）组件对100名PD患者和79名对照受试者的EEG信号进行分类，这些数据来自三个公共静息状态EEG数据集。提出了一种系统的可解释性方法，包括嵌入可视化和基于mha的时间和光谱分析，通过统计测试来提高鉴别生物标志物的识别能力。在一个大型的、多中心的受试者队列中，实验结果证明了该框架在感兴趣的频率间隔中检测有意义的脑电图模式的能力。可解释性分析表明，MHA集中于输入信号中特定的时间斑块，这与分类结果相关。频谱分析发现Theta和Beta波段的显著功率差异，捕捉PD患者认知和运动功能障碍的神经模式。此外，注意引导分割提高了光谱生物标志物的灵敏度，如Alpha/Theta比和Beta相对功率，与先前的文献一致。此外，该方法在Theta上的平均auc最高为0.91±0.01，Alpha上为0.84±0.03，All-band上为0.81±0.01，达到了最先进的性能，同时也证明了对异构数据的鲁棒性。

{"title":"Parkinson’s disease detection from resting-state EEG using a Transformer model with Multi-Head Attention explanations","authors":"Valeriana Mancazzo , Elena Sibilano , Domenico Buongiorno, Raffaele Carli, Vitoantonio Bevilacqua, Antonio Brunetti","doi":"10.1016/j.bspc.2026.109811","DOIUrl":"10.1016/j.bspc.2026.109811","url":null,"abstract":"<div><div>Despite the proven potential of using Deep Learning (DL) models based on electroencephalographic (EEG) signals to detect neurological disorders like Parkinson’s Disease (PD), their adoption in clinical practice is limited due to insufficient reliability and generalizability. We propose an interpretable end-to-end DL framework leveraging the Multi-Head Attention (MHA) component of Transformers to classify EEG signals of 100 PD patients and 79 control subjects across three public resting-state EEG datasets. A systematic interpretability approach, including embedding visualization and MHA-based temporal and spectral analysis through statistical tests, is proposed to enhance the identification of discriminative biomarkers. Experimental findings across a large, multi-centric cohort of subjects demonstrated the framework’s capability to detect meaningful EEG patterns in the frequency intervals of interest.</div><div>Interpretability analysis revealed that MHA focused on specific temporal patches in the input signal, which correlated to the classification outcomes. Spectral analysis identified significant power differences in Theta and Beta bands, capturing neural patterns of cognitive and motor dysfunction in PD. Furthermore, attention-guided segmentation improved the sensitivity of spectral biomarkers, such as Alpha/Theta ratio and Beta relative power, consistent with prior literature. Moreover, the proposed approach yielded the highest epoch-level mean AUCs of <span><math><mrow><mn>0</mn><mo>.</mo><mn>91</mn><mspace></mspace><mo>±</mo><mspace></mspace><mn>0</mn><mo>.</mo><mn>01</mn></mrow></math></span> on Theta, <span><math><mrow><mn>0</mn><mo>.</mo><mn>84</mn><mspace></mspace><mo>±</mo><mspace></mspace><mn>0</mn><mo>.</mo><mn>03</mn></mrow></math></span> on Alpha, and <span><math><mrow><mn>0</mn><mo>.</mo><mn>81</mn><mspace></mspace><mo>±</mo><mspace></mspace><mn>0</mn><mo>.</mo><mn>01</mn></mrow></math></span> on All-band, achieving state-of-the-art performances while also demonstrating robustness to heterogeneous data.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109811"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191956","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

Integrating EEG microstates and functional connectivity via cross-attention for emotion recognition in virtual reality 基于交叉注意的脑电微态与功能连接集成在虚拟现实中的情感识别

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-06 DOI: 10.1016/j.bspc.2026.109685

Yicai Bai , Yucheng Zhou , Jinqi Dong , Dengjiujiu He , Chao Jiang , Jinglu Hu , Yingjie Li

Emotion is fundamental to human cognition and behavior, and electroencephalography (EEG), with its high temporal resolution, provides a powerful approach to investigate the neural activity underlying emotional processing. We combined EEG with Virtual Reality technology to conduct an EEG-based emotion experiment in a more immersive environment. Moreover, EEG microstate and functional connectivity features are closely related yet capturing their complex nonlinear interactions remains challenging. To address this challenge, we proposed a deep learning framework that integrates Cross-Attention mechanisms with convolutional neural networks (CNN) to model these interactions. Specifically, Cross-Attention captures inter-feature dependencies, while CNN performs hierarchical feature extraction. Experimental results demonstrate that our framework significantly outperforms the baseline CNN model, particularly in recognizing subtle emotional states such as neutral emotion. Notably, this improvement is driven by the synergistic interaction between microstate and functional connectivity features, thereby improving model interpretability. These findings highlight the potential of Cross-Attention CNN to elucidate the complex nonlinear neural mechanisms underlying emotional processing.

情绪是人类认知和行为的基础，脑电图（EEG）以其高时间分辨率为研究情绪加工背后的神经活动提供了有力的手段。我们将脑电图与虚拟现实技术相结合，在更加身临其境的环境中进行基于脑电图的情绪实验。此外，脑电微观状态和功能连接特征密切相关，但捕捉它们之间复杂的非线性相互作用仍然是一个挑战。为了应对这一挑战，我们提出了一个深度学习框架，该框架将交叉注意机制与卷积神经网络（CNN）集成在一起，以模拟这些相互作用。具体来说，交叉注意捕获特征之间的依赖关系，而CNN执行分层特征提取。实验结果表明，我们的框架显著优于基线CNN模型，特别是在识别微妙的情绪状态（如中性情绪）方面。值得注意的是，这种改进是由微状态和功能连接特征之间的协同交互驱动的，从而提高了模型的可解释性。这些发现强调了交叉注意CNN在阐明情绪处理背后复杂的非线性神经机制方面的潜力。

{"title":"Integrating EEG microstates and functional connectivity via cross-attention for emotion recognition in virtual reality","authors":"Yicai Bai , Yucheng Zhou , Jinqi Dong , Dengjiujiu He , Chao Jiang , Jinglu Hu , Yingjie Li","doi":"10.1016/j.bspc.2026.109685","DOIUrl":"10.1016/j.bspc.2026.109685","url":null,"abstract":"<div><div>Emotion is fundamental to human cognition and behavior, and electroencephalography (EEG), with its high temporal resolution, provides a powerful approach to investigate the neural activity underlying emotional processing. We combined EEG with Virtual Reality technology to conduct an EEG-based emotion experiment in a more immersive environment. Moreover, EEG microstate and functional connectivity features are closely related yet capturing their complex nonlinear interactions remains challenging. To address this challenge, we proposed a deep learning framework that integrates Cross-Attention mechanisms with convolutional neural networks (CNN) to model these interactions. Specifically, Cross-Attention captures inter-feature dependencies, while CNN performs hierarchical feature extraction. Experimental results demonstrate that our framework significantly outperforms the baseline CNN model, particularly in recognizing subtle emotional states such as neutral emotion. Notably, this improvement is driven by the synergistic interaction between microstate and functional connectivity features, thereby improving model interpretability. These findings highlight the potential of Cross-Attention CNN to elucidate the complex nonlinear neural mechanisms underlying emotional processing.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109685"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146193098","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

Assessing equivalence in raw accelerometer outputs across different brands using shaker table validation: a comparative analysis with filtering and linear regression techniques 评估等效的原始加速度计输出跨不同品牌使用振动台验证：与过滤和线性回归技术的比较分析

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-12 DOI: 10.1016/j.bspc.2026.109627

Hannah J. Coyle-Asbil , Bernadette Murphy , Lori Ann Vallis

This study had two major aims, first to determine whether the raw acceleration data from ActiGraph versus non-ActiGraph accelerometers were equivalent and, second to apply and compare different equivalency approaches: 1) linear regression, 2) lowpass 15 Hz, 3) lowpass 20 Hz, and 4) lowpass 30 Hz filters. ActiGraph GT3X+ (n = 8), ActiGraph wGT3X-BT (n = 10), ActiGraph GT9X (n = 8; primary and GT9X2), OPAL (n = 6) and GENEActiv (n = 5) accelerometers were affixed to a multi-axis shaker table and sinusoidal oscillations were introduced, spanning the entire dynamic range (±0.005 G to ± 8 G). The averages of the trials were compared according to the different techniques. Linear regression models were fitted to align the non-ActiGraph to the ActiGraph data, and lowpass filters were applied to the non-Actigraph data. Equivalency was assessed using two one sided t-tests of equivalence. The results indicated either statistically insignificant or negligible mean differences at the lower frequency oscillations, however at higher oscillations the recordings were significantly distinct and vastly varied across the different techniques. For example, the mean difference between the wGT3X-BT − GA at trial 28 unprocessed was −2788.09 mg, −394.93 mg for the linear regression, 1358.56 mg for the 15 Hz filter, and 182.94 mg for both the 20 Hz and 30 Hz filters. In conclusion, there are minor differences between ActiGraph and non-ActiGraph accelerometers at low frequencies; however, at high frequencies, 20 Hz low-pass filters are effective in improving equivalency. Our findings provide insight into device equivalency and important guidance for researchers to consider when harmonizing data across accelerometer devices.

这项研究有两个主要目的，首先是确定来自ActiGraph和非ActiGraph加速度计的原始加速度数据是否等效，其次是应用和比较不同的等效方法：1)线性回归，2)低通15hz， 3)低通20hz和4)低通30hz滤波器。将ActiGraph GT3X+ (n = 8)、ActiGraph wGT3X-BT （n = 10）、ActiGraph GT9X （n = 8； primary和GT9X2）、OPAL （n = 6）和GENEActiv （n = 5）加速度计固定在多轴激振台上，引入正弦振荡，整个动态范围（±0.005 G至±8g）。根据不同的技术比较试验的平均值。拟合线性回归模型使非ActiGraph数据与ActiGraph数据对齐，并对非ActiGraph数据应用低通滤波器。使用等效性的两个单侧t检验来评估等效性。结果表明，在较低的频率振荡中，统计上不显著或可以忽略不计的平均差异，然而在较高的振荡中，不同技术的记录显着不同且差异很大。例如，试验28未处理的wgt3g - bt - GA之间的平均差异为- 2788.09 mg，线性回归为- 394.93 mg， 15 Hz滤波器为1358.56 mg， 20 Hz和30 Hz滤波器为182.94 mg。总之，在低频下，ActiGraph和非ActiGraph加速度计之间存在微小差异；然而，在高频下，20hz低通滤波器可以有效地提高等效性。我们的研究结果为研究人员在协调加速度计设备之间的数据时提供了对设备等效性的见解和重要指导。

{"title":"Assessing equivalence in raw accelerometer outputs across different brands using shaker table validation: a comparative analysis with filtering and linear regression techniques","authors":"Hannah J. Coyle-Asbil , Bernadette Murphy , Lori Ann Vallis","doi":"10.1016/j.bspc.2026.109627","DOIUrl":"10.1016/j.bspc.2026.109627","url":null,"abstract":"<div><div>This study had two major aims, first to determine whether the raw acceleration data from ActiGraph versus non-ActiGraph accelerometers were equivalent and, second to apply and compare different equivalency approaches: 1) linear regression, 2) lowpass 15 Hz, 3) lowpass 20 Hz, and 4) lowpass 30 Hz filters. ActiGraph GT3X+ (n = 8), ActiGraph wGT3X-BT (n = 10), ActiGraph GT9X (n = 8; primary and GT9X2), OPAL (n = 6) and GENEActiv (n = 5) accelerometers were affixed to a multi-axis shaker table and sinusoidal oscillations were introduced, spanning the entire dynamic range (±0.005 G to ± 8 G). The averages of the trials were compared according to the different techniques. Linear regression models were fitted to align the non-ActiGraph to the ActiGraph data, and lowpass filters were applied to the non-Actigraph data. Equivalency was assessed using two one sided t-tests of equivalence. The results indicated either statistically insignificant or negligible mean differences at the lower frequency oscillations, however at higher oscillations the recordings were significantly distinct and vastly varied across the different techniques. For example, the mean difference between the wGT3X-BT − GA at trial 28 unprocessed was −2788.09 m<em>g</em>, −394.93 m<em>g</em> for the linear regression, 1358.56 m<em>g</em> for the 15 Hz filter, and 182.94 m<em>g</em> for both the 20 Hz and 30 Hz filters. In conclusion, there are minor differences between ActiGraph and non-ActiGraph accelerometers at low frequencies; however, at high frequencies, 20 Hz low-pass filters are effective in improving equivalency. Our findings provide insight into device equivalency and important guidance for researchers to consider when harmonizing data across accelerometer devices.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109627"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146193094","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

AMU-Net: A hybrid U-net model with ASPP and mamba gating mechanism for coronary artery segmentation AMU-Net：一个具有ASPP和曼巴门控机制的混合U-net模型用于冠状动脉分割

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-14 DOI: 10.1016/j.bspc.2026.109837

Hengjie Ouyang , Yang Li , Yunlong Zhao , Benqiang Yang , Libo Zhang

Automatic coronary artery segmentation is crucial for diagnosing cardiovascular diseases but remains a challenging task. Key challenges in CTA images include low contrast and artifacts, which hinder the identification of small vessels and the preservation of topological connectivity. A successful model must therefore effectively capture both global context for overall structure and fine local details for precise boundaries. To address these multifaceted challenges, this paper proposes AMU-Net, an integrated model that combines three key components. First, a novel MambaASPP module in the bottleneck layer effectively captures both global context and local details. Second, a composite loss function (CE_DC_CBDC) specifically targets small vessels and preserves topological connectivity. Third, deep supervision ensures robust feature learning throughout the network. Our evaluation assesses both segmentation accuracy and connectivity. Experimental results show that AMU-Net significantly outperforms existing models across metrics for both accuracy (IoU, Dice) and connectivity (clDice, cbDice), demonstrating its practical value. As shown in our ablation study, this synergistic design achieves an improvement of approximately 3% in the Dice similarity coefficient over a standard U-Net baseline.

冠状动脉自动分割是诊断心血管疾病的关键，但仍然是一项具有挑战性的任务。CTA图像的主要挑战包括低对比度和伪影，这阻碍了小血管的识别和拓扑连通性的保存。因此，一个成功的模型必须有效地捕捉整体结构的全局背景和精确边界的精细局部细节。为了应对这些多方面的挑战，本文提出了AMU-Net，这是一个结合了三个关键组件的集成模型。首先，瓶颈层中的一个新的MambaASPP模块可以有效地捕获全局上下文和本地细节。其次，复合损失函数（CE_DC_CBDC）专门针对小血管并保持拓扑连通性。第三，深度监督确保了整个网络的鲁棒性特征学习。我们的评估评估了分割的准确性和连通性。实验结果表明，AMU-Net在精度（IoU, Dice）和连通性（clDice, cbDice）方面都明显优于现有模型，证明了其实用价值。正如我们的消融研究所示，这种协同设计在Dice相似系数方面比标准U-Net基线提高了约3%。

{"title":"AMU-Net: A hybrid U-net model with ASPP and mamba gating mechanism for coronary artery segmentation","authors":"Hengjie Ouyang , Yang Li , Yunlong Zhao , Benqiang Yang , Libo Zhang","doi":"10.1016/j.bspc.2026.109837","DOIUrl":"10.1016/j.bspc.2026.109837","url":null,"abstract":"<div><div>Automatic coronary artery segmentation is crucial for diagnosing cardiovascular diseases but remains a challenging task. Key challenges in CTA images include low contrast and artifacts, which hinder the identification of small vessels and the preservation of topological connectivity. A successful model must therefore effectively capture both global context for overall structure and fine local details for precise boundaries. To address these multifaceted challenges, this paper proposes AMU-Net, an integrated model that combines three key components. First, a novel MambaASPP module in the bottleneck layer effectively captures both global context and local details. Second, a composite loss function (CE_DC_CBDC) specifically targets small vessels and preserves topological connectivity. Third, deep supervision ensures robust feature learning throughout the network. Our evaluation assesses both segmentation accuracy and connectivity. Experimental results show that AMU-Net significantly outperforms existing models across metrics for both accuracy (IoU, Dice) and connectivity (clDice, cbDice), demonstrating its practical value. As shown in our ablation study, this synergistic design achieves an improvement of approximately 3% in the Dice similarity coefficient over a standard U-Net baseline.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109837"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146193106","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

SWDL: Stratum-Wise Difference Learning with deep Laplacian pyramid for semi-supervised 3D intracranial hemorrhage segmentation 基于深度拉普拉斯金字塔的分层差分学习半监督三维颅内出血分割

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-14 DOI: 10.1016/j.bspc.2026.109744

Cheng Wang , Siqi Chen , Donghua Mi , Jingyang Zhang , Yang Chen , Yudong Zhang , Yinsheng Li

Recent advances in medical imaging have established deep learning-based segmentation as the predominant approach, although it typically requires large amounts of manually annotated data. However, obtaining annotations for intracranial hemorrhage (ICH) remains particularly challenging due to the tedious and costly labeling process. Semi-supervised learning (SSL) has emerged as a promising solution to address the scarcity of labeled data, especially in volumetric medical image segmentation. Unlike conventional SSL methods that primarily focus on high-confidence pseudo-labels or consistency regularization, we propose SWDL-Net, a novel SSL framework that exploits the complementary advantages of Laplacian pyramid and deep convolutional upsampling. The Laplacian pyramid excels at edge sharpening, while deep convolutions enhance detail precision through flexible feature mapping. Our framework achieves superior segmentation of lesion details and boundaries through a difference learning mechanism that effectively integrates these complementary approaches. Extensive experiments on a 271-case ICH dataset and public benchmarks demonstrate that SWDL-Net outperforms current state-of-the-art methods in scenarios with only 2% labeled data. Additional evaluations on the publicly available Brain Hemorrhage Segmentation Dataset (BHSD) with 5% labeled data further confirm the superiority of our approach. Code and data have been released at https://github.com/SIAT-CT-LAB/SWDL.

医学成像的最新进展已经建立了基于深度学习的分割作为主要方法，尽管它通常需要大量手动注释的数据。然而，由于繁琐和昂贵的标记过程，获得颅内出血（ICH）的注释仍然特别具有挑战性。半监督学习（SSL）已成为解决标记数据稀缺性的一种有前途的解决方案，特别是在体积医学图像分割中。与主要关注高置信度伪标签或一致性正则化的传统SSL方法不同，我们提出了SWDL-Net，这是一种利用拉普拉斯金字塔和深度卷积上采样互补优势的新型SSL框架。拉普拉斯金字塔擅长边缘锐化，而深度卷积通过灵活的特征映射提高细节精度。我们的框架通过有效整合这些互补方法的差异学习机制，实现了对病变细节和边界的优越分割。在271例ICH数据集和公共基准测试上进行的大量实验表明，在只有2%标记数据的情况下，SWDL-Net优于当前最先进的方法。对公开可用的脑出血分割数据集（BHSD）进行了额外的评估，其中有5%的标记数据进一步证实了我们方法的优越性。代码和数据已在https://github.com/SIAT-CT-LAB/SWDL上发布。

{"title":"SWDL: Stratum-Wise Difference Learning with deep Laplacian pyramid for semi-supervised 3D intracranial hemorrhage segmentation","authors":"Cheng Wang , Siqi Chen , Donghua Mi , Jingyang Zhang , Yang Chen , Yudong Zhang , Yinsheng Li","doi":"10.1016/j.bspc.2026.109744","DOIUrl":"10.1016/j.bspc.2026.109744","url":null,"abstract":"<div><div>Recent advances in medical imaging have established deep learning-based segmentation as the predominant approach, although it typically requires large amounts of manually annotated data. However, obtaining annotations for intracranial hemorrhage (ICH) remains particularly challenging due to the tedious and costly labeling process. Semi-supervised learning (SSL) has emerged as a promising solution to address the scarcity of labeled data, especially in volumetric medical image segmentation. Unlike conventional SSL methods that primarily focus on high-confidence pseudo-labels or consistency regularization, we propose SWDL-Net, a novel SSL framework that exploits the complementary advantages of Laplacian pyramid and deep convolutional upsampling. The Laplacian pyramid excels at edge sharpening, while deep convolutions enhance detail precision through flexible feature mapping. Our framework achieves superior segmentation of lesion details and boundaries through a difference learning mechanism that effectively integrates these complementary approaches. Extensive experiments on a 271-case ICH dataset and public benchmarks demonstrate that SWDL-Net outperforms current state-of-the-art methods in scenarios with only 2% labeled data. Additional evaluations on the publicly available Brain Hemorrhage Segmentation Dataset (BHSD) with 5% labeled data further confirm the superiority of our approach. Code and data have been released at <span><span>https://github.com/SIAT-CT-LAB/SWDL</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109744"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146193103","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Optimized deep learning framework for periodontal disease severity prediction and treatment recommendation 优化牙周病严重程度预测和治疗建议的深度学习框架

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-06-15 Epub Date: 2026-02-13 DOI: 10.1016/j.bspc.2026.109768

R. Kausalya , J. Anitha Ruth

Periodontal disease severity prediction is essential for early detection and personalized therapy planning for patients. Early identification allows for prompt actions that have the potential of significantly enhancing the results for patients and reduce the progression of the disease. This study proposes a novel deep learning (DL) model to accurately classify periodontal disease severity. The model integrates advanced preprocessing techniques, feature extraction, periodontal severity prediction and treatment recommendation based on detected severity to ensure enhanced performance. The preprocessing phase employs YOLOv7, image scaling, augmentation, Contrast Limited Adaptive Histogram Equalization (CLAHE), and bilateral filtering to improve image clarity and quality. In the feature extraction phase, an improved ResNet50 model is applied to obtain discriminative features from the preprocessed output.The prediction of periodontal disease severity is achieved using an Optimal Attention CNN-based Bidirectional Gated Recurrent Unit (ATT-CNN-BiGRU) model, which effectively captures sequential dependencies for accurate classification into severity levels such as normal, mild, moderate, and severe. The Optimal ATT-CNN-BiGRU model is fine-tuned with the White Shark Optimizer (WSO). Depending on the predicted severity level, the model provides tailored treatment suggestions, including professional cleaning, scaling, root planning, antibiotic therapy, and surgical intervention for more severe cases, thereby enhancing prediction accuracy.The model, implemented using the MATLAB platform and assessed using a Kaggle dataset, obtained a remarkable 98.91% accuracy rate and demonstrated strong performance metrics. The results demonstrate that combining advanced techniques with optimization strategies significantly enhances the accuracy and durability of the model in detecting periodontal disease severity.

牙周病严重程度预测对于患者的早期发现和个性化治疗计划至关重要。早期发现有助于迅速采取行动，有可能显著提高患者的治疗效果，并减少疾病的进展。本研究提出一种新的深度学习（DL）模型来准确分类牙周病的严重程度。该模型集成了先进的预处理技术、特征提取、牙周严重程度预测和基于检测到的严重程度的治疗建议，以确保提高性能。预处理阶段采用YOLOv7、图像缩放、增强、对比度有限自适应直方图均衡化（CLAHE）和双边滤波来提高图像清晰度和质量。在特征提取阶段，采用改进的ResNet50模型从预处理后的输出中获得判别特征。牙周病严重程度的预测是使用基于优化注意力cnn的双向门控复发单元（ATT-CNN-BiGRU）模型实现的，该模型有效地捕获顺序依赖关系，以准确分类为严重级别，如正常，轻度，中度和严重。最优的ATT-CNN-BiGRU模型是用白鲨优化器（WSO）微调的。根据预测的严重程度，该模型提供量身定制的治疗建议，包括专业清洁、刮治、牙根规划、抗生素治疗，并对更严重的病例进行手术干预，从而提高了预测的准确性。该模型在MATLAB平台上实现，并使用Kaggle数据集进行评估，获得了98.91%的准确率，并展示了强大的性能指标。结果表明，将先进技术与优化策略相结合，可以显著提高模型检测牙周病严重程度的准确性和耐久性。

{"title":"Optimized deep learning framework for periodontal disease severity prediction and treatment recommendation","authors":"R. Kausalya , J. Anitha Ruth","doi":"10.1016/j.bspc.2026.109768","DOIUrl":"10.1016/j.bspc.2026.109768","url":null,"abstract":"<div><div>Periodontal disease severity prediction is essential for early detection and personalized therapy planning for patients. Early identification allows for prompt actions that have the potential of significantly enhancing the results for patients and reduce the progression of the disease. This study proposes a novel deep learning (DL) model to accurately classify periodontal disease severity. The model integrates advanced preprocessing techniques, feature extraction, periodontal severity prediction and treatment recommendation based on detected severity to ensure enhanced performance. The preprocessing phase employs YOLOv7, image scaling, augmentation, Contrast Limited Adaptive Histogram Equalization (CLAHE), and bilateral filtering to improve image clarity and quality. In the feature extraction phase, an improved ResNet50 model is applied to obtain discriminative features from the preprocessed output.The prediction of periodontal disease severity is achieved using an Optimal Attention CNN-based Bidirectional Gated Recurrent Unit (ATT-CNN-BiGRU) model, which effectively captures sequential dependencies for accurate classification into severity levels such as normal, mild, moderate, and severe. The Optimal ATT-CNN-BiGRU model is fine-tuned with the White Shark Optimizer (WSO). Depending on the predicted severity level, the model provides tailored treatment suggestions, including professional cleaning, scaling, root planning, antibiotic therapy, and surgical intervention for more severe cases, thereby enhancing prediction accuracy.The model, implemented using the MATLAB platform and assessed using a Kaggle dataset, obtained a remarkable 98.91% accuracy rate and demonstrated strong performance metrics. The results demonstrate that combining advanced techniques with optimization strategies significantly enhances the accuracy and durability of the model in detecting periodontal disease severity.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"119 ","pages":"Article 109768"},"PeriodicalIF":4.9,"publicationDate":"2026-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192563","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