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

DRHAFHNet: Dense resolution high-order attention forward harmonic network-based learning effectiveness of shopfloor employees with digital twin DRHAFHNet：基于数字孪生的车间员工密集分辨率高阶注意前向谐波网络的学习效果

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109605

Chandra Mohini C. P., V. Raghavendran

The quick advances in the field of self-driving vehicles and connected automobiles have increased the commercial worth of automobile applications. Digital Twin is employed as a promising technology to modernize the automotive industry. Moreover, the development of digital twins has offered smart manufacturing systems with knowledge-making capabilities. Hence, the training in a virtual environment minimizes the errors on the shop floor. Still, the extraction of relevant insights to establish the optimal course sequence for the shop floor employees is computationally difficult. To overcome such issues, this paper develops the Dense Resolution High-order Attention Forward Harmonic Network (DRHAFHNet)-based course sequence recommendation for learning the effectiveness of shopfloor employees with a digital twin. The shop floor owner collects the data from the physical space, and the cloud server stores the data from the shop floor owner. The twin manager collects the data from the cloud server and simulates it in the virtual space. The virtual data is stored in the cloud, and the course sequence recommendation is performed by the DRHAFHNet using a digital twin E-learning platform. Moreover, the proposed model attains the Normalized Mean Square Error (MSE), Normalized Mean Absolute Percentage Error (MAPE), Normalized Root MSE (RMSE), Normalized Mean Absolute Percentage (MAP), and latency of 0.334, 0.332, 0.342, 0.299, and 4.99 ms.

自动驾驶汽车和网联汽车领域的快速发展增加了汽车应用的商业价值。数字孪生技术是一项很有前途的汽车工业现代化技术。此外，数字孪生体的发展提供了具有知识制造能力的智能制造系统。因此，在虚拟环境中的培训可以最大限度地减少车间的错误。然而，提取相关的见解以建立车间员工的最佳课程顺序在计算上是困难的。为了克服这些问题，本文开发了基于密集分辨率高阶注意前向谐波网络（DRHAFHNet）的课程顺序推荐，用于学习车间员工的数字孪生的有效性。车间所有者从物理空间收集数据，云服务器存储车间所有者的数据。twin管理器从云服务器收集数据，并在虚拟空间中进行模拟。虚拟数据存储在云中，课程顺序推荐由DRHAFHNet使用数字孪生电子学习平台执行。此外，该模型还实现了归一化均方误差（MSE）、归一化平均绝对百分比误差（MAPE）、归一化根误差（RMSE）、归一化平均绝对百分比误差（MAP）和延迟分别为0.334、0.332、0.342、0.299和4.99 ms。

{"title":"DRHAFHNet: Dense resolution high-order attention forward harmonic network-based learning effectiveness of shopfloor employees with digital twin","authors":"Chandra Mohini C. P., V. Raghavendran","doi":"10.1016/j.bspc.2026.109605","DOIUrl":"10.1016/j.bspc.2026.109605","url":null,"abstract":"<div><div>The quick advances in the field of self-driving vehicles and connected automobiles have increased the commercial worth of automobile applications. Digital Twin is employed as a promising technology to modernize the automotive industry. Moreover, the development of digital twins has offered smart manufacturing systems with knowledge-making capabilities. Hence, the training in a virtual environment minimizes the errors on the shop floor. Still, the extraction of relevant insights to establish the optimal course sequence for the shop floor employees is computationally difficult. To overcome such issues, this paper develops the Dense Resolution High-order Attention Forward Harmonic Network (DRHAFHNet)-based course sequence recommendation for learning the effectiveness of shopfloor employees with a digital twin. The shop floor owner collects the data from the physical space, and the cloud server stores the data from the shop floor owner. The twin manager collects the data from the cloud server and simulates it in the virtual space. The virtual data is stored in the cloud, and the course sequence recommendation is performed by the DRHAFHNet using a digital twin E-learning platform. Moreover, the proposed model attains the Normalized Mean Square Error (MSE), Normalized Mean Absolute Percentage Error (MAPE), Normalized Root MSE (RMSE), Normalized Mean Absolute Percentage (MAP), and latency of 0.334, 0.332, 0.342, 0.299, and 4.99 ms.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109605"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039006","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

Multistage transfer learning for skin squamous cell carcinoma histopathology image classification 多阶段迁移学习在皮肤鳞状细胞癌组织病理图像分类中的应用

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109660

Gelan Ayana , Beshatu Debela Wako , So-yun Park , Kwangcheol Casey Jeong , Soon‑Do Yoon , Se‑woon Choe

Squamous cell carcinoma (SCC) is the second most common form of skin cancer with significant public health implications owing to its potential for metastasis if not detected and treated early. Traditional diagnostic methods, which rely on histopathological analysis, face challenges, such as variability in tissue morphology and dependence on expert interpretation, leading to inconsistent diagnoses. To address these issues, this work proposes a novel multistage transfer learning (MSTL) approach that leverages deep learning models for automated SCC diagnosis from histopathological images. The MSTL framework begins with a model pretrained on the extensive ImageNet dataset, fine-tuned on a large breast histopathology dataset to capture domain-specific features, and further refined on a smaller SCC histopathology dataset. Vision transformer (ViT) models have been employed, marking a pioneering application in SCC analysis. The experimental results showed that the MSTL-based ViT model achieved state-of-the-art accuracy of 0.9752, precision of 0.9708, recall of 0.9734, F1 score of 0.9741, and an area under receiver operating curve (AUC) of 0.9739, thereby setting a new benchmark. Furthermore, the MSTL approach demonstrated superior training efficiency, with reduced loss and faster convergence compared to the conventional transfer learning models, without excessive computational costs. Evaluation on an independent dataset confirmed the robustness of the MSTL approach with the ViT model, achieving an AUC of 0.9437. The MSTL approach also exhibited strong transferability, with high Pearson correlation coefficients between the transferability measures and AUCs. Further investigations are needed to assess the generalizability of MSTL to other cancers and its applicability in clinical settings.

鳞状细胞癌（SCC）是第二种最常见的皮肤癌，如果不及早发现和治疗，可能会发生转移，对公众健康造成重大影响。传统的诊断方法依赖于组织病理学分析，面临着挑战，如组织形态的可变性和对专家解释的依赖，导致诊断不一致。为了解决这些问题，本研究提出了一种新的多阶段迁移学习（MSTL）方法，该方法利用深度学习模型从组织病理学图像中自动诊断SCC。MSTL框架首先在广泛的ImageNet数据集上对模型进行预训练，在大型乳腺组织病理学数据集上进行微调，以捕获特定领域的特征，并在较小的SCC组织病理学数据集上进一步完善。视觉变压器（Vision transformer, ViT）模型在SCC分析中具有开创性的应用。实验结果表明，基于mstl的ViT模型的准确率为0.9752，精密度为0.9708，召回率为0.9734，F1得分为0.9741，接收者工作曲线下面积（AUC）为0.9739，达到了一个新的基准。此外，与传统迁移学习模型相比，MSTL方法具有更低的损失和更快的收敛速度，并且没有过多的计算成本，显示出更高的训练效率。在独立数据集上的评估证实了MSTL方法与ViT模型的鲁棒性，AUC为0.9437。MSTL方法也表现出很强的可转移性，可转移性指标与auc之间具有较高的Pearson相关系数。需要进一步的研究来评估MSTL在其他癌症中的普遍性及其在临床环境中的适用性。

{"title":"Multistage transfer learning for skin squamous cell carcinoma histopathology image classification","authors":"Gelan Ayana , Beshatu Debela Wako , So-yun Park , Kwangcheol Casey Jeong , Soon‑Do Yoon , Se‑woon Choe","doi":"10.1016/j.bspc.2026.109660","DOIUrl":"10.1016/j.bspc.2026.109660","url":null,"abstract":"<div><div>Squamous cell carcinoma (SCC) is the second most common form of skin cancer with significant public health implications owing to its potential for metastasis if not detected and treated early. Traditional diagnostic methods, which rely on histopathological analysis, face challenges, such as variability in tissue morphology and dependence on expert interpretation, leading to inconsistent diagnoses. To address these issues, this work proposes a novel multistage transfer learning (MSTL) approach that leverages deep learning models for automated SCC diagnosis from histopathological images. The MSTL framework begins with a model pretrained on the extensive ImageNet dataset, fine-tuned on a large breast histopathology dataset to capture domain-specific features, and further refined on a smaller SCC histopathology dataset. Vision transformer (ViT) models have been employed, marking a pioneering application in SCC analysis. The experimental results showed that the MSTL-based ViT model achieved state-of-the-art accuracy of 0.9752, precision of 0.9708, recall of 0.9734, F1 score of 0.9741, and an area under receiver operating curve (AUC) of 0.9739, thereby setting a new benchmark. Furthermore, the MSTL approach demonstrated superior training efficiency, with reduced loss and faster convergence compared to the conventional transfer learning models, without excessive computational costs. Evaluation on an independent dataset confirmed the robustness of the MSTL approach with the ViT model, achieving an AUC of 0.9437. The MSTL approach also exhibited strong transferability, with high Pearson correlation coefficients between the transferability measures and AUCs. Further investigations are needed to assess the generalizability of MSTL to other cancers and its applicability in clinical settings.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109660"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039008","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

Heterogeneous multi-network cross pseudo-supervised medical image segmentation 异构多网络交叉伪监督医学图像分割

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109641

Li Kang, Chuanghong Zhao, Jianjun Huang, Zhixin Gong

Medical image segmentation is a crucial component of medical image processing. However, most medical image segmentation methods require a large number of labeled samples to train neural networks. The annotation work for medical image segmentation is both labor-intensive and technically demanding, resulting in high costs for obtaining high-quality annotations. This paper aims to perform medical image segmentation under conditions of data lacking annotations using semi-supervised techniques. This paper proposes a multi-network cross pseudo-supervision method that leverages CNN and Transformer networks. By utilizing the differences between these networks, the method allows different networks to learn from the various perspectives provided by others. To further enhance these differences and discover common features among pixels of the same category in different samples, a contrastive learning method is integrated into one of the CNN networks. Experiments conducted on benchmark datasets have validated the effectiveness of the proposed method across different label proportions. Ablation studies on multi-network cross pseudo-supervision and contrastive loss demonstrate the effectiveness of the network architecture and contrastive learning. Extensive experimental results show that the semi-supervised proposed in this paper is superior to the existing six semi-supervised learning methods. With a small number of annotated samples, this method can significantly improve network performance. This advancement significantly enhances the accuracy and efficiency of medical image segmentation, crucial for precise lesion detection, treatment planning, and therapeutic effect assessment, even with limited annotated data.

医学图像分割是医学图像处理的重要组成部分。然而，大多数医学图像分割方法需要大量的标记样本来训练神经网络。医学图像分割的标注工作劳动强度大，技术要求高，获取高质量标注的成本高。本文旨在利用半监督技术对缺乏标注的数据进行医学图像分割。本文提出了一种利用CNN和Transformer网络的多网络交叉伪监督方法。通过利用这些网络之间的差异，该方法允许不同的网络从其他网络提供的不同角度进行学习。为了进一步增强这些差异，并发现不同样本中同一类别像素之间的共同特征，我们将对比学习方法集成到其中一个CNN网络中。在基准数据集上进行的实验验证了该方法在不同标签比例下的有效性。对多网络交叉伪监督和对比损失的研究证明了网络结构和对比学习的有效性。大量的实验结果表明，本文提出的半监督学习方法优于现有的六种半监督学习方法。在标注样本数量较少的情况下，该方法可以显著提高网络性能。这一进步显著提高了医学图像分割的准确性和效率，对于精确的病变检测、治疗计划和治疗效果评估至关重要，即使在有限的注释数据下也是如此。

{"title":"Heterogeneous multi-network cross pseudo-supervised medical image segmentation","authors":"Li Kang, Chuanghong Zhao, Jianjun Huang, Zhixin Gong","doi":"10.1016/j.bspc.2026.109641","DOIUrl":"10.1016/j.bspc.2026.109641","url":null,"abstract":"<div><div>Medical image segmentation is a crucial component of medical image processing. However, most medical image segmentation methods require a large number of labeled samples to train neural networks. The annotation work for medical image segmentation is both labor-intensive and technically demanding, resulting in high costs for obtaining high-quality annotations. This paper aims to perform medical image segmentation under conditions of data lacking annotations using semi-supervised techniques. This paper proposes a multi-network cross pseudo-supervision method that leverages CNN and Transformer networks. By utilizing the differences between these networks, the method allows different networks to learn from the various perspectives provided by others. To further enhance these differences and discover common features among pixels of the same category in different samples, a contrastive learning method is integrated into one of the CNN networks. Experiments conducted on benchmark datasets have validated the effectiveness of the proposed method across different label proportions. Ablation studies on multi-network cross pseudo-supervision and contrastive loss demonstrate the effectiveness of the network architecture and contrastive learning. Extensive experimental results show that the semi-supervised proposed in this paper is superior to the existing six semi-supervised learning methods. With a small number of annotated samples, this method can significantly improve network performance. This advancement significantly enhances the accuracy and efficiency of medical image segmentation, crucial for precise lesion detection, treatment planning, and therapeutic effect assessment, even with limited annotated data.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109641"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038914","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

Multi-Source Unsupervised Domain Adaptation with dual alignment for medical image segmentation 基于双对齐的多源无监督域自适应医学图像分割

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109678

Chunping Gao , Lihua Guo , Qi Wu

In clinical practice, physicians usually review medical images from various modalities to obtain comprehensive structural insights that support accurate diagnosis. Inspired by this practice, Multi-Source Unsupervised Domain Adaptation (MUDA) aims to improve model generalization on an unlabeled target domain by leveraging structural knowledge from multiple labeled source domains. In medical image segmentation, existing MUDA frameworks mainly employ adversarial training to achieve domain adaptation, but they often capture limited structural information. In this work, we propose a dual alignment MUDA framework for medical image segmentation, which jointly uses image alignment and feature alignment to facilitate sufficient knowledge transfer. For image alignment, we apply Fourier-based domain adaptation (FDA) to mitigate appearance discrepancies between the source and target domains. For feature alignment, we integrate adversarial learning and curvature-based boundary consistency constraints to align global predictions and preserve local boundary details. Furthermore, we develop a performance-aware ensemble strategy that adaptively emphasizes models with superior performance, thereby improving prediction robustness. Extensive experiments on three publicly available datasets demonstrate that our method significantly outperforms existing state-of-the-art approaches. The proposed method enables robust unsupervised domain adaptation among medical images from multiple domains with large domain shifts.

在临床实践中，医生通常会检查各种形式的医学图像，以获得全面的结构见解，从而支持准确的诊断。受这一实践的启发，多源无监督域自适应（Multi-Source Unsupervised Domain Adaptation， MUDA）旨在利用来自多个标记源域的结构知识来提高模型在未标记目标域上的泛化。在医学图像分割中，现有的MUDA框架主要采用对抗训练来实现领域自适应，但往往捕获有限的结构信息。在这项工作中，我们提出了一种双对齐的医学图像分割MUDA框架，该框架联合使用图像对齐和特征对齐来促进充分的知识转移。对于图像对齐，我们应用基于傅立叶的域适应（FDA）来减轻源域和目标域之间的外观差异。对于特征对齐，我们集成了对抗学习和基于曲率的边界一致性约束来对齐全局预测并保留局部边界细节。此外，我们开发了一种性能感知集成策略，该策略自适应地强调具有优异性能的模型，从而提高了预测的鲁棒性。在三个公开可用的数据集上进行的大量实验表明，我们的方法明显优于现有的最先进的方法。该方法实现了多域大位移医学图像的鲁棒无监督域自适应。

{"title":"Multi-Source Unsupervised Domain Adaptation with dual alignment for medical image segmentation","authors":"Chunping Gao , Lihua Guo , Qi Wu","doi":"10.1016/j.bspc.2026.109678","DOIUrl":"10.1016/j.bspc.2026.109678","url":null,"abstract":"<div><div>In clinical practice, physicians usually review medical images from various modalities to obtain comprehensive structural insights that support accurate diagnosis. Inspired by this practice, Multi-Source Unsupervised Domain Adaptation (MUDA) aims to improve model generalization on an unlabeled target domain by leveraging structural knowledge from multiple labeled source domains. In medical image segmentation, existing MUDA frameworks mainly employ adversarial training to achieve domain adaptation, but they often capture limited structural information. In this work, we propose a dual alignment MUDA framework for medical image segmentation, which jointly uses image alignment and feature alignment to facilitate sufficient knowledge transfer. For image alignment, we apply Fourier-based domain adaptation (FDA) to mitigate appearance discrepancies between the source and target domains. For feature alignment, we integrate adversarial learning and curvature-based boundary consistency constraints to align global predictions and preserve local boundary details. Furthermore, we develop a performance-aware ensemble strategy that adaptively emphasizes models with superior performance, thereby improving prediction robustness. Extensive experiments on three publicly available datasets demonstrate that our method significantly outperforms existing state-of-the-art approaches. The proposed method enables robust unsupervised domain adaptation among medical images from multiple domains with large domain shifts.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109678"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039010","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

Topology-convolutional long short-term memory network for epileptic seizure prediction: An interpretable deep learning framework based on a multi-center dataset 用于癫痫发作预测的拓扑卷积长短期记忆网络：基于多中心数据集的可解释深度学习框架

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109652

Jina E. , Lin Wang , Bingchao Wan , Wenjie Yu , Yushi Chen , Lingxia Fei , Feng Yang , Jun Zhuang

Epilepsy is a common chronic neurological disorder, and its recurrent seizures severely impact patients’ physical and mental health as well as social functioning. This study presents a novel deep learning framework inspired by neuroscience, utilizing a Topological Convolutional Neural Network with Long Short-Term Memory (TSCNN-LSTM). This framework aims to enhance the quality of life for epilepsy patients and optimize clinical management strategies. The model utilizes a convolutional neural network to extract time–frequency features from EEG signals and an LSTM network to simulate the dynamic interactions between brain regions, achieving accurate seizure prediction. The TSCNN-LSTM model was evaluated across three datasets: CHB-MIT, Siena, and our clinical dataset. On CHB-MIT, the model achieved 96.0% sensitivity, 95.0% specificity, and 98.0% AUC. Siena dataset validation yielded 89.2% accuracy and 94.8% AUC. Clinical data evaluation demonstrated 85.9% accuracy and 92.5% AUC, confirming robust performance across diverse recording conditions. These consistent metrics across heterogeneous datasets validate the model’s exceptional predictive accuracy and cross-dataset generalizability, establishing its potential for clinical seizure prediction applications.

癫痫是一种常见的慢性神经系统疾病，其反复发作严重影响患者的身心健康和社会功能。本研究提出了一种受神经科学启发的新型深度学习框架，利用具有长短期记忆的拓扑卷积神经网络（TSCNN-LSTM）。该框架旨在提高癫痫患者的生活质量，优化临床管理策略。该模型利用卷积神经网络提取脑电图信号的时频特征，利用LSTM网络模拟脑区之间的动态相互作用，实现准确的癫痫发作预测。TSCNN-LSTM模型在三个数据集上进行评估：CHB-MIT、Siena和我们的临床数据集。在CHB-MIT上，模型的灵敏度为96.0%，特异性为95.0%，AUC为98.0%。Siena数据集验证的准确率为89.2%，AUC为94.8%。临床数据评估显示准确率为85.9%，AUC为92.5%，证实了在不同记录条件下的稳健性能。这些跨异构数据集的一致指标验证了该模型卓越的预测准确性和跨数据集的通用性，从而确立了其在临床癫痫发作预测应用中的潜力。

{"title":"Topology-convolutional long short-term memory network for epileptic seizure prediction: An interpretable deep learning framework based on a multi-center dataset","authors":"Jina E. , Lin Wang , Bingchao Wan , Wenjie Yu , Yushi Chen , Lingxia Fei , Feng Yang , Jun Zhuang","doi":"10.1016/j.bspc.2026.109652","DOIUrl":"10.1016/j.bspc.2026.109652","url":null,"abstract":"<div><div>Epilepsy is a common chronic neurological disorder, and its recurrent seizures severely impact patients’ physical and mental health as well as social functioning. This study presents a novel deep learning framework inspired by neuroscience, utilizing a Topological Convolutional Neural Network with Long Short-Term Memory (TSCNN-LSTM). This framework aims to enhance the quality of life for epilepsy patients and optimize clinical management strategies. The model utilizes a convolutional neural network to extract time–frequency features from EEG signals and an LSTM network to simulate the dynamic interactions between brain regions, achieving accurate seizure prediction. The TSCNN-LSTM model was evaluated across three datasets: CHB-MIT, Siena, and our clinical dataset. On CHB-MIT, the model achieved 96.0% sensitivity, 95.0% specificity, and 98.0% AUC. Siena dataset validation yielded 89.2% accuracy and 94.8% AUC. Clinical data evaluation demonstrated 85.9% accuracy and 92.5% AUC, confirming robust performance across diverse recording conditions. These consistent metrics across heterogeneous datasets validate the model’s exceptional predictive accuracy and cross-dataset generalizability, establishing its potential for clinical seizure prediction applications.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109652"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039016","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

An enhanced educational competition optimizer and feedforward neural networks for automatic seizure detection in EEG signals 一种用于脑电图信号癫痫自动检测的增强教育竞争优化器和前馈神经网络

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109656

Asmaa Hammad , Mosa E. Hosney , Marwa M. Emam , Nagwan Abdel Samee , Reem Ibrahim Alkanhel , Essam H. Houssein

Accurately detecting and staging epileptic seizures is crucial in medical diagnostics, as early identification significantly improves patient survival rates. Electroencephalogram (EEG) signals are widely used for seizure detection, yet no universally accepted feature set exists for this purpose. While incorporating all possible EEG features may enhance classification accuracy, excessive dimensionality introduces redundancy and inefficiency, ultimately reducing overall performance. To overcome this challenge, we propose a novel wrapper-based feature selection method, ECO-OL, which integrates Orthogonal Learning (OL) with the Educational Competition Optimizer (ECO) to form a hybrid optimization algorithm. ECO-OL enhances the search process, selects the most relevant features, and prevents premature convergence while maintaining classifier performance. Additionally, we introduce a hybrid classification model, ECO-OL with MLPNN, which combines the improved ECO-OL feature selection approach with a multi-layer perceptron neural network (MLPNN) to optimize EEG seizure classification. ECO-OL was thoroughly assessed with the CEC’22 test suite and the Khas EEG seizure detection dataset, consisting of preictal, interictal, and ictal EEG signals. The method was applied to classify interictal vs. ictal, interictal vs. preictal, preictal vs. ictal, and all three classes combined. Experimental results demonstrate that ECO-OL outperforms ECO and six widely used metaheuristic algorithms in diversity, convergence, and statistical measures. The proposed method achieves 99.2% accuracy in ictal–preictal classification and 95.9% in interictal–preictal classification, surpassing existing methods by 1.2% and 15.8%, respectively. This study provides a robust computational framework for seizure detection, with promising applications in AI-driven medical research and bioinformatics.

在医学诊断中，准确检测和分期癫痫发作是至关重要的，因为早期识别可以显著提高患者的存活率。脑电图（EEG）信号被广泛用于癫痫发作检测，但没有普遍接受的特征集存在。虽然结合所有可能的EEG特征可以提高分类精度，但过多的维度会引入冗余和低效率，最终降低整体性能。为了克服这一挑战，我们提出了一种新的基于包装的特征选择方法ECO-OL，该方法将正交学习（OL）与教育竞争优化器（ECO）相结合，形成一种混合优化算法。ECO-OL增强了搜索过程，选择最相关的特征，并在保持分类器性能的同时防止过早收敛。此外，我们还引入了一种混合分类模型ECO-OL与MLPNN，该模型将改进的ECO-OL特征选择方法与多层感知器神经网络（MLPNN）相结合，以优化脑电图发作分类。ECO-OL通过CEC ' 22测试套件和Khas脑电图发作检测数据集进行全面评估，该数据集包括癫痫发作前、癫痫发作间和癫痫发作时的脑电图信号。该方法被应用于间隔与间隔、间隔与间隔、间隔与间隔、间隔与间隔，以及所有三个类别的组合。实验结果表明，ECO- ol在多样性、收敛性和统计度量方面优于ECO和六种广泛使用的元启发式算法。该方法在统计-预测分类中准确率达到99.2%，在统计-预测分类中准确率达到95.9%，分别比现有方法高1.2%和15.8%。这项研究为癫痫检测提供了一个强大的计算框架，在人工智能驱动的医学研究和生物信息学中具有广阔的应用前景。

{"title":"An enhanced educational competition optimizer and feedforward neural networks for automatic seizure detection in EEG signals","authors":"Asmaa Hammad , Mosa E. Hosney , Marwa M. Emam , Nagwan Abdel Samee , Reem Ibrahim Alkanhel , Essam H. Houssein","doi":"10.1016/j.bspc.2026.109656","DOIUrl":"10.1016/j.bspc.2026.109656","url":null,"abstract":"<div><div>Accurately detecting and staging epileptic seizures is crucial in medical diagnostics, as early identification significantly improves patient survival rates. Electroencephalogram (EEG) signals are widely used for seizure detection, yet no universally accepted feature set exists for this purpose. While incorporating all possible EEG features may enhance classification accuracy, excessive dimensionality introduces redundancy and inefficiency, ultimately reducing overall performance. To overcome this challenge, we propose a novel wrapper-based feature selection method, ECO-OL, which integrates Orthogonal Learning (OL) with the Educational Competition Optimizer (ECO) to form a hybrid optimization algorithm. ECO-OL enhances the search process, selects the most relevant features, and prevents premature convergence while maintaining classifier performance. Additionally, we introduce a hybrid classification model, ECO-OL with MLPNN, which combines the improved ECO-OL feature selection approach with a multi-layer perceptron neural network (MLPNN) to optimize EEG seizure classification. ECO-OL was thoroughly assessed with the CEC’22 test suite and the Khas EEG seizure detection dataset, consisting of preictal, interictal, and ictal EEG signals. The method was applied to classify interictal vs. ictal, interictal vs. preictal, preictal vs. ictal, and all three classes combined. Experimental results demonstrate that ECO-OL outperforms ECO and six widely used metaheuristic algorithms in diversity, convergence, and statistical measures. The proposed method achieves 99.2% accuracy in ictal–preictal classification and 95.9% in interictal–preictal classification, surpassing existing methods by 1.2% and 15.8%, respectively. This study provides a robust computational framework for seizure detection, with promising applications in AI-driven medical research and bioinformatics.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109656"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038913","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

MCAB-GFEResNet: A multimodal fusion model for pre-treatment prediction of neoadjuvant chemoradiotherapy response in rectal cancer MCAB-GFEResNet：用于直肠癌新辅助放化疗治疗前反应预测的多模式融合模型

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109672

Qing Lu , Longbo Zheng , Jionglong Su , Weimin Ma , Hui Ma , Yulin Zhang

An innovative multimodal deep learning model, MCAB-GFEResNet, was developed to predict treatment response to neoadjuvant chemoradiotherapy (nCRT) in patients with locally advanced rectal cancer. The system integrates multiparametric baseline data acquired prior to nCRT initiation: (1) pre-treatment whole-slide biopsy images (WSI) obtained during initial diagnostic workup, (2) baseline magnetic resonance imaging (MRI) performed after biopsy but before nCRT commencement, and (3) pretreatment clinical biomarkers (including carcinoembryonic antigen [CEA] levels) collected contemporaneously with imaging. The architecture innovatively introduces two key components: a Multimodal Clue Attention Bridge (MCAB) fusion strategy that achieves deep feature fusion through cross-modal interaction, and a Global Feature Enhancement (GFE) module that precisely captures discriminative features in treatment response via dual-dimensional (channel and spatial) feature aggregation mechanisms. Prospective validation on 185 patients achieved 97.21% accuracy, a 5.6–20.2% absolute improvement over reference fusion methods, with 97.25% sensitivity using only pretreatment data. This enables therapeutic decisions to be made at least 4–6 weeks earlier compared to post-surgical pathology-dependent approaches. It demonstrates significant advantages in both predictive accuracy and temporal efficiency, establishing the first multimodal framework for precision prediction of nCRT response using solely pretreatment data in rectal cancer.

一种创新的多模式深度学习模型MCAB-GFEResNet被开发出来，用于预测局部晚期直肠癌患者对新辅助放化疗（nCRT）的治疗反应。该系统集成了在nCRT开始之前获得的多参数基线数据：(1)在初始诊断检查期间获得的预处理全切片活检图像（WSI），(2)在活检后但在nCRT开始之前进行的基线磁共振成像（MRI），以及(3)与成像同时收集的预处理临床生物标志物（包括癌胚抗原[CEA]水平）。该架构创新地引入了两个关键组件：通过跨模态交互实现深度特征融合的多模态线索注意桥（MCAB）融合策略，以及通过二维（通道和空间）特征聚合机制精确捕获治疗响应中判别特征的全局特征增强（GFE）模块。185例患者的前瞻性验证准确率为97.21%，比参考融合方法绝对提高5.6-20.2%，仅使用预处理数据的敏感性为97.25%。与术后病理依赖的方法相比，这使得治疗决定至少提前4-6周做出。它在预测准确性和时间效率方面都显示出显著的优势，建立了第一个多模式框架，用于仅使用直肠癌预处理数据精确预测nCRT反应。

{"title":"MCAB-GFEResNet: A multimodal fusion model for pre-treatment prediction of neoadjuvant chemoradiotherapy response in rectal cancer","authors":"Qing Lu , Longbo Zheng , Jionglong Su , Weimin Ma , Hui Ma , Yulin Zhang","doi":"10.1016/j.bspc.2026.109672","DOIUrl":"10.1016/j.bspc.2026.109672","url":null,"abstract":"<div><div>An innovative multimodal deep learning model, MCAB-GFEResNet, was developed to predict treatment response to neoadjuvant chemoradiotherapy (nCRT) in patients with locally advanced rectal cancer. The system integrates multiparametric baseline data acquired prior to nCRT initiation: (1) pre-treatment whole-slide biopsy images (WSI) obtained during initial diagnostic workup, (2) baseline magnetic resonance imaging (MRI) performed after biopsy but before nCRT commencement, and (3) pretreatment clinical biomarkers (including carcinoembryonic antigen [CEA] levels) collected contemporaneously with imaging. The architecture innovatively introduces two key components: a Multimodal Clue Attention Bridge (MCAB) fusion strategy that achieves deep feature fusion through cross-modal interaction, and a Global Feature Enhancement (GFE) module that precisely captures discriminative features in treatment response via dual-dimensional (channel and spatial) feature aggregation mechanisms. Prospective validation on 185 patients achieved 97.21% accuracy, a 5.6–20.2% absolute improvement over reference fusion methods, with 97.25% sensitivity using only pretreatment data. This enables therapeutic decisions to be made at least 4–6 weeks earlier compared to post-surgical pathology-dependent approaches. It demonstrates significant advantages in both predictive accuracy and temporal efficiency, establishing the first multimodal framework for precision prediction of nCRT response using solely pretreatment data in rectal cancer.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109672"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038916","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

Deep learning ensemble approach to multi-wavelength photoplethysmogram for sleep stage detection 用于睡眠阶段检测的多波长光容积图深度学习集成方法

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109689

Yuna Park , Heeseung Cho , Junhyoung Oh

Sleep stages are vital for assessing sleep quality, and vary significantly between individuals. Polysomnography (PSG), the clinical gold standard, is impractical for continuous monitoring due to its complexity. Wearable devices using photoplethysmography (PPG) offer a scalable, non-invasive alternative, but PPG signals are prone to motion artifacts. To address this, we utilized multiwavelength PPG signals (two green, red, and infrared channels) from a commercial smartwatch, adhering to American Academy of Sleep Medicine (AASM) standards, to classify sleep stages. We compared machine learning and deep learning models, evaluating single-channel and multi-channel fusion strategies. Our convolutional neural network combined with the gated recurrent unit (CNN+GRU) late ensemble fusion model achieved the highest performance, with a precision of 73.8% and a Cohen kappa of 0.535, outperforming the best single green channel model by 2.8% (

p = 0.0015

, rank-biserial correlation = 0.6670). Green channels excelled in detecting rapid eye movement stages, while the red channel enhanced the differentiation between deep sleep (DS) and light sleep (LS) by 10%–14%. Temporal Saliency Rescaling (TSR) analysis confirmed the focus of the model on relevant signal regions, improving robustness against motion artifacts. Wilcoxon signed rank tests validated the superiority of the multichannel model (

p = 0.0015

for green channel 1,

p = 0.0147

for green channel 2). These findings highlight the potential of multiwavelength PPG for accurate, interpretable sleep stage classification, allowing scalable sleep monitoring as a viable alternative to PSG.

睡眠阶段对评估睡眠质量至关重要，而且因人而异。多导睡眠图（PSG）作为临床金标准，由于其复杂性，不适合连续监测。使用光电容积脉搏波（PPG）的可穿戴设备提供了一种可扩展的、非侵入性的替代方案，但PPG信号容易产生运动伪影。为了解决这个问题，我们使用了来自商用智能手表的多波长PPG信号（两个绿色，红色和红外通道），遵循美国睡眠医学学会（AASM）的标准，对睡眠阶段进行分类。我们比较了机器学习和深度学习模型，评估了单通道和多通道融合策略。我们的卷积神经网络结合门控循环单元（CNN+GRU）晚集成融合模型取得了最高的性能，精度为73.8%，Cohen kappa为0.535，比最佳单一绿色通道模型高出2.8% （p=0.0015，秩双序列相关性= 0.6670）。绿色通道在检测快速眼动阶段方面表现出色，而红色通道使深度睡眠（DS）和浅睡眠（LS）之间的区分增强了10%-14%。时间显著性重新缩放（TSR）分析证实了模型对相关信号区域的关注，提高了对运动伪像的鲁棒性。Wilcoxon符号秩检验验证了多通道模型的优越性（绿色通道1的p=0.0015，绿色通道2的p=0.0147）。这些发现突出了多波长PPG在准确、可解释的睡眠阶段分类方面的潜力，使可扩展的睡眠监测成为PSG的可行替代方案。

{"title":"Deep learning ensemble approach to multi-wavelength photoplethysmogram for sleep stage detection","authors":"Yuna Park , Heeseung Cho , Junhyoung Oh","doi":"10.1016/j.bspc.2026.109689","DOIUrl":"10.1016/j.bspc.2026.109689","url":null,"abstract":"<div><div>Sleep stages are vital for assessing sleep quality, and vary significantly between individuals. Polysomnography (PSG), the clinical gold standard, is impractical for continuous monitoring due to its complexity. Wearable devices using photoplethysmography (PPG) offer a scalable, non-invasive alternative, but PPG signals are prone to motion artifacts. To address this, we utilized multiwavelength PPG signals (two green, red, and infrared channels) from a commercial smartwatch, adhering to American Academy of Sleep Medicine (AASM) standards, to classify sleep stages. We compared machine learning and deep learning models, evaluating single-channel and multi-channel fusion strategies. Our convolutional neural network combined with the gated recurrent unit (CNN+GRU) late ensemble fusion model achieved the highest performance, with a precision of 73.8% and a Cohen kappa of 0.535, outperforming the best single green channel model by 2.8% (<span><math><mrow><mi>p</mi><mo>=</mo><mn>0</mn><mo>.</mo><mn>0015</mn></mrow></math></span>, rank-biserial correlation = 0.6670). Green channels excelled in detecting rapid eye movement stages, while the red channel enhanced the differentiation between deep sleep (DS) and light sleep (LS) by 10%–14%. Temporal Saliency Rescaling (TSR) analysis confirmed the focus of the model on relevant signal regions, improving robustness against motion artifacts. Wilcoxon signed rank tests validated the superiority of the multichannel model (<span><math><mrow><mi>p</mi><mo>=</mo><mn>0</mn><mo>.</mo><mn>0015</mn></mrow></math></span> for green channel 1, <span><math><mrow><mi>p</mi><mo>=</mo><mn>0</mn><mo>.</mo><mn>0147</mn></mrow></math></span> for green channel 2). These findings highlight the potential of multiwavelength PPG for accurate, interpretable sleep stage classification, allowing scalable sleep monitoring as a viable alternative to PSG.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109689"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039011","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

Innovative Algorithm for Keratoconus Intelligent Grading Using Variational Encoding Bayesian Gaussian Mixture Model 基于变分编码贝叶斯-高斯混合模型的圆锥角膜智能评分创新算法

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109642

Zuoping Tan , Xuan Chen , Shuangcheng Li , Lijuan Yue , Tinghui Huang , Rui Yao , Jianying Lv , Jing Li , Caiye Fan , Riwei Wang , Yuanyuan Wang , Yan Wang

Background

The variational encoding Bayesian Gaussian mixture model is a novel unsupervised machine learning model that combines variational encoding and Bayesian inference. The model uses a variational encoder to learn the continuous, structured latent space of a dataset and combines multiple Gaussian mixture distributions, offering robust learning and generalisation capabilities. Although it is well-suited for modelling complex relationships, this model has not yet been explored in grading keratoconus, a blinding eye disease with unclear etiology or standardised diagnostic treatment systems. Therefore, this study applies the Bayesian Gaussian mixture model to categorise keratoconus severity and identifies relatively sensitive features for early diagnoses and interventions, potentially improving clinical decision-making and visual outcomes.

Results

The eyes of 456 patients with keratoconus are analysed. Using the variational self-encoder Bayesian Gaussian mixture model for unsupervised keratoconus grading, with classification into four categories, an accuracy of 84% is achieved, over 30% higher than that of other unsupervised algorithms, such as K-means and DBSCAN. The area under the curve values of the four categories are 0.923, 0.856, 0.789, and 0.986. Moreover, further analyses show that features such as minimum thickness (Pachy min) are more sensitive to grading outcomes.

Conclusions

The variational encoded Bayesian Gaussian model effectively captures the key features of keratoconus severity and enables accurate, automatic grading. This provides valuable clinical references for assessing disease progression and treatment, and presents a novel approach for the early diagnosis of keratoconus.

变分编码贝叶斯高斯混合模型是一种将变分编码与贝叶斯推理相结合的新型无监督机器学习模型。该模型使用变分编码器来学习数据集的连续、结构化潜在空间，并结合多个高斯混合分布，提供强大的学习和泛化能力。虽然它非常适合建模复杂的关系，但该模型尚未在圆锥角膜分级中进行探索，圆锥角膜是一种病因不明或标准化诊断治疗系统的致盲眼病。因此，本研究应用贝叶斯-高斯混合模型对圆锥角膜的严重程度进行分类，并识别相对敏感的特征，用于早期诊断和干预，可能改善临床决策和视力结果。结果对456例圆锥角膜患者的眼部进行了分析。使用变分自编码器贝叶斯高斯混合模型进行无监督圆锥角膜评分，将其分为四类，准确率达到84%，比其他无监督算法（如K-means和DBSCAN）提高30%以上。四类曲线下面积分别为0.923、0.856、0.789、0.986。此外，进一步的分析表明，最小厚度（Pachy min）等特征对分级结果更为敏感。结论变分编码贝叶斯高斯模型能有效地捕捉圆锥角膜严重程度的关键特征，实现准确、自动的分级。这为评估疾病进展和治疗提供了有价值的临床参考，并为圆锥角膜的早期诊断提供了一种新的方法。

{"title":"Innovative Algorithm for Keratoconus Intelligent Grading Using Variational Encoding Bayesian Gaussian Mixture Model","authors":"Zuoping Tan , Xuan Chen , Shuangcheng Li , Lijuan Yue , Tinghui Huang , Rui Yao , Jianying Lv , Jing Li , Caiye Fan , Riwei Wang , Yuanyuan Wang , Yan Wang","doi":"10.1016/j.bspc.2026.109642","DOIUrl":"10.1016/j.bspc.2026.109642","url":null,"abstract":"<div><h3>Background</h3><div>The variational encoding Bayesian Gaussian mixture model is a novel unsupervised machine learning model that combines variational encoding and Bayesian inference. The model uses a variational encoder to learn the continuous, structured latent space of a dataset and combines multiple Gaussian mixture distributions, offering robust learning and generalisation capabilities. Although it is well-suited for modelling complex relationships, this model has not yet been explored in grading keratoconus, a blinding eye disease with unclear etiology or standardised diagnostic treatment systems. Therefore, this study applies the Bayesian Gaussian mixture model to categorise keratoconus severity and identifies relatively sensitive features for early diagnoses and interventions, potentially improving clinical decision-making and visual outcomes.</div></div><div><h3>Results</h3><div>The eyes of 456 patients with keratoconus are analysed. Using the variational self-encoder Bayesian Gaussian mixture model for unsupervised keratoconus grading, with classification into four categories, an accuracy of 84% is achieved, over 30% higher than that of other unsupervised algorithms, such as K-means and DBSCAN. The area under the curve values of the four categories are 0.923, 0.856, 0.789, and 0.986. Moreover, further analyses show that features such as minimum thickness (Pachy min) are more sensitive to grading outcomes.</div></div><div><h3>Conclusions</h3><div>The variational encoded Bayesian Gaussian model effectively captures the key features of keratoconus severity and enables accurate, automatic grading. This provides valuable clinical references for assessing disease progression and treatment, and presents a novel approach for the early diagnosis of keratoconus.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109642"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039007","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A novel multimodal learning method for predicting treatment resistance in MPO-AAV with lung involvement 一种新的多模态学习方法预测MPO-AAV累及肺部的治疗耐药

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control

Pub Date : 2026-01-23 DOI: 10.1016/j.bspc.2026.109643

Yinan Zhang , Yigang Pei , Ting Meng , Yu Wang , Yong Zhong , Yixiong Liang

The prediction of treatment resistance in myeloperoxidase (MPO)-anti-neutrophil cytoplasmic antibody (ANCA)-associated vasculitis (MPO-AAV) with lung involvement is critical, as a significant number of patients develop resistance to existing therapies, resulting in more severe consequences. Previous methods leverage radiomics analysis to extract features from CT images, which are then combined with clinical information for prediction. However, these methods rely on radiologists to manually delineate regions of interest (ROIs) for CT images, a process that is time-consuming and labor-intensive, and their performance is constrained by a simple linear fusion of radiomic features and clinical information. In this paper, we propose a novel multimodal learning method for predicting treatment resistance in MPO-AAV with lung involvement. We first introduce the lesion-aware re-embedding (LARE) module and the cross-slice interaction (CSI) module to adapt the powerful vision foundation model (VFM) for extracting visual representation from CT images, thereby eliminating the dependence on lesion ROIs during inference. Furthermore, we utilize Transformer to extract high-level semantic information from raw clinical data and incorporate a learnable multimodal feature fusion module (MFFM) to enhance the interaction and integration of multimodal features. We construct experiments on a dual-center dataset. The experimental results show that our proposed method achieves superior performance, exceeding previous methods by a clear margin. Additionally, we visualize the perception map of lesion ROIs generated by the LARE module and assess the importance of clinical attributes, qualitatively analyzing the performance of the proposed method. The code and trained models are publicly available at https://github.com/CVIU-CSU/PTRNet.

髓过氧化物酶(MPO)-抗中性粒细胞胞浆抗体(ANCA)-相关血管炎（MPO- aav）伴肺受损伤的治疗耐药预测至关重要，因为大量患者对现有治疗产生耐药性，从而导致更严重的后果。以前的方法利用放射组学分析从CT图像中提取特征，然后将其与临床信息相结合进行预测。然而，这些方法依赖于放射科医生手动划定CT图像的感兴趣区域（roi），这一过程既耗时又费力，而且它们的性能受到放射学特征和临床信息的简单线性融合的限制。在本文中，我们提出了一种新的多模态学习方法来预测MPO-AAV累及肺部的治疗耐药性。我们首先引入病变感知重嵌入（LARE）模块和横切面交互（CSI）模块，采用强大的视觉基础模型（VFM）从CT图像中提取视觉表示，从而消除了推理过程中对病变roi的依赖。此外，我们利用Transformer从原始临床数据中提取高级语义信息，并结合可学习的多模态特征融合模块（MFFM）来增强多模态特征的交互和集成。我们在双中心数据集上构建实验。实验结果表明，该方法取得了较好的性能，明显优于以往的方法。此外，我们可视化了由LARE模块生成的病变roi的感知图，并评估临床属性的重要性，定性分析了所提出方法的性能。代码和经过训练的模型可在https://github.com/CVIU-CSU/PTRNet上公开获得。

{"title":"A novel multimodal learning method for predicting treatment resistance in MPO-AAV with lung involvement","authors":"Yinan Zhang , Yigang Pei , Ting Meng , Yu Wang , Yong Zhong , Yixiong Liang","doi":"10.1016/j.bspc.2026.109643","DOIUrl":"10.1016/j.bspc.2026.109643","url":null,"abstract":"<div><div>The prediction of treatment resistance in myeloperoxidase (MPO)-anti-neutrophil cytoplasmic antibody (ANCA)-associated vasculitis (MPO-AAV) with lung involvement is critical, as a significant number of patients develop resistance to existing therapies, resulting in more severe consequences. Previous methods leverage radiomics analysis to extract features from CT images, which are then combined with clinical information for prediction. However, these methods rely on radiologists to manually delineate regions of interest (ROIs) for CT images, a process that is time-consuming and labor-intensive, and their performance is constrained by a simple linear fusion of radiomic features and clinical information. In this paper, we propose a novel multimodal learning method for predicting treatment resistance in MPO-AAV with lung involvement. We first introduce the lesion-aware re-embedding (LARE) module and the cross-slice interaction (CSI) module to adapt the powerful vision foundation model (VFM) for extracting visual representation from CT images, thereby eliminating the dependence on lesion ROIs during inference. Furthermore, we utilize Transformer to extract high-level semantic information from raw clinical data and incorporate a learnable multimodal feature fusion module (MFFM) to enhance the interaction and integration of multimodal features. We construct experiments on a dual-center dataset. The experimental results show that our proposed method achieves superior performance, exceeding previous methods by a clear margin. Additionally, we visualize the perception map of lesion ROIs generated by the LARE module and assess the importance of clinical attributes, qualitatively analyzing the performance of the proposed method. The code and trained models are publicly available at <span><span>https://github.com/CVIU-CSU/PTRNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"117 ","pages":"Article 109643"},"PeriodicalIF":4.9,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039015","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