首页 > 最新文献
paperpass

Applied Soft Computing最新文献

英文 中文
Boundary-guided large-scale vision model for unified multi-domain industrial anomaly detection 统一多领域工业异常检测的边界引导大规模视觉模型
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-01-13 DOI: 10.1016/j.asoc.2026.114650
Zixuan Zhang , Fan Shi , Chen Jia , Mianzhao Wang , Xu Cheng
Extracting shared boundary cues across different anomaly domains is critical to enhancing generalization on unseen data, thereby laying the foundation for unified industrial anomaly detection paradigms. Existing unified detection paradigms often directly extract discriminative features from multiple data domains. However, due to the inherent semantic gaps between different data sources, bridging this disparity within a shared feature representation across multiple data domains remains a key challenge. To address this challenge, we propose a boundary-guided large-scale vision model that extracts commonalities across diverse domains. Specifically, we generate initial feature embeddings by establishing a multi-domain normal sample repository and employing a parameter coupling strategy. This captures shared boundary information across different data domains, thereby reducing the inherent semantic gaps. For anomalous feature synthesis, we incorporate this boundary information into the generation process, ensuring that the synthesized features retain critical structural details while expanding the coverage of potential anomalous data distributions. Additionally, to enhance feature space separation between normal and anomalous samples, we introduce a hybrid constraint optimization mechanism that improves the discriminative ability of the model. Extensive experiments on the MVTec AD, VisA, and MPDD datasets demonstrate that our method achieves state-of-the-art performance across various industrial scenarios. Experimental results demonstrate the effectiveness of boundary-guided shared information for multi-domain anomaly detection.
提取跨不同异常域的共享边界线索对于增强对未见数据的泛化至关重要,从而为统一的工业异常检测范式奠定基础。现有的统一检测范式往往直接从多个数据域中提取判别特征。然而,由于不同数据源之间固有的语义差距,在跨多个数据域的共享特征表示中弥合这种差距仍然是一个关键的挑战。为了解决这一挑战,我们提出了一个边界引导的大规模视觉模型,该模型可以提取不同领域的共性。具体来说,我们通过建立多域正常样本库和采用参数耦合策略来生成初始特征嵌入。这捕获了跨不同数据域的共享边界信息,从而减少了固有的语义差距。对于异常特征合成,我们将这些边界信息纳入到生成过程中,确保合成的特征保留关键的结构细节,同时扩大潜在异常数据分布的覆盖范围。此外,为了增强正常和异常样本之间的特征空间分离,我们引入了一种混合约束优化机制,提高了模型的判别能力。在MVTec AD、VisA和MPDD数据集上进行的大量实验表明,我们的方法可以在各种工业场景中实现最先进的性能。实验结果证明了边界引导共享信息用于多域异常检测的有效性。
{"title":"Boundary-guided large-scale vision model for unified multi-domain industrial anomaly detection","authors":"Zixuan Zhang ,&nbsp;Fan Shi ,&nbsp;Chen Jia ,&nbsp;Mianzhao Wang ,&nbsp;Xu Cheng","doi":"10.1016/j.asoc.2026.114650","DOIUrl":"10.1016/j.asoc.2026.114650","url":null,"abstract":"<div><div>Extracting shared boundary cues across different anomaly domains is critical to enhancing generalization on unseen data, thereby laying the foundation for unified industrial anomaly detection paradigms. Existing unified detection paradigms often directly extract discriminative features from multiple data domains. However, due to the inherent semantic gaps between different data sources, bridging this disparity within a shared feature representation across multiple data domains remains a key challenge. To address this challenge, we propose a boundary-guided large-scale vision model that extracts commonalities across diverse domains. Specifically, we generate initial feature embeddings by establishing a multi-domain normal sample repository and employing a parameter coupling strategy. This captures shared boundary information across different data domains, thereby reducing the inherent semantic gaps. For anomalous feature synthesis, we incorporate this boundary information into the generation process, ensuring that the synthesized features retain critical structural details while expanding the coverage of potential anomalous data distributions. Additionally, to enhance feature space separation between normal and anomalous samples, we introduce a hybrid constraint optimization mechanism that improves the discriminative ability of the model. Extensive experiments on the MVTec AD, VisA, and MPDD datasets demonstrate that our method achieves state-of-the-art performance across various industrial scenarios. Experimental results demonstrate the effectiveness of boundary-guided shared information for multi-domain anomaly detection.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114650"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039851","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
End-to-end discrete cosine transform integration in spectral convolutional neural networks for resource-efficient deep learning 频谱卷积神经网络中资源高效深度学习的端到端离散余弦变换积分
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-01-13 DOI: 10.1016/j.asoc.2026.114599
Ibrahim Yousef Alshareef , Ab Al-Hadi Ab Rahman , Nuzhat Khan , Hasan Alqaraghuli
Spectral convolutional neural networks using Fast Fourier Transform (FFT) often suffer from high computational complexity and memory demands due to complex valued operations and the need for inverse transforms limiting their deployment on resource constrained devices. This paper presents a novel end-to-end spectral convolutional neural network (SpCNN) architecture that operates entirely in the Discrete Cosine Transform (DCT) domain, eliminating the need for inverse transformations and complex arithmetic. Leveraging the DCT’s real valued representation and superior energy compaction, the proposed design significantly reduces computational workload and memory usage while preserving classification accuracy. Key innovations include the removal of IFFT layers, a frequency domain adaptive activation function (FReLU), and a DCT optimized spectral pooling mechanism, each tailored for deployment in low power, resource constrained environments. Experimental evaluations on MNIST and a 94-class ASCII dataset demonstrate the model’s efficiency: LeNet5-DCT achieves a 37.96% FLOPs reduction, 18.45% lower memory usage, and 96.56% test accuracy, while VGG7-DCT achieves a 33.95% FLOPs reduction, 14.32% lower memory usage, and 90.62% test accuracy. The architecture also shows strong robustness to quantization, confirming its suitability for edge AI applications and low energy inference. This work provides a scalable, hardware efficient spectral learning framework, paving the way for future hybrid spectral models optimized for embedded environments.
使用快速傅里叶变换(FFT)的频谱卷积神经网络通常由于复杂的值运算和对逆变换的需求而遭受高计算复杂度和内存需求的困扰,这限制了它们在资源受限设备上的部署。本文提出了一种全新的端到端频谱卷积神经网络(SpCNN)架构,该架构完全在离散余弦变换(DCT)域中运行,消除了对逆变换和复杂算法的需要。利用DCT的真实值表示和优越的能量压缩,提出的设计显着减少了计算工作量和内存使用,同时保持了分类准确性。关键创新包括去除IFFT层、频域自适应激活函数(FReLU)和DCT优化的频谱池机制,每个创新都是针对低功耗、资源受限环境的部署而定制的。在MNIST和94类ASCII数据集上的实验评估证明了该模型的效率:LeNet5-DCT实现了37.96%的FLOPs降低,18.45%的内存使用降低,96.56%的测试准确率,而VGG7-DCT实现了33.95%的FLOPs降低,14.32%的内存使用降低,90.62%的测试准确率。该架构还显示出强大的量化鲁棒性,证实了其适用于边缘人工智能应用和低能量推理。这项工作提供了一个可扩展的、硬件高效的光谱学习框架,为未来针对嵌入式环境优化的混合光谱模型铺平了道路。
{"title":"End-to-end discrete cosine transform integration in spectral convolutional neural networks for resource-efficient deep learning","authors":"Ibrahim Yousef Alshareef ,&nbsp;Ab Al-Hadi Ab Rahman ,&nbsp;Nuzhat Khan ,&nbsp;Hasan Alqaraghuli","doi":"10.1016/j.asoc.2026.114599","DOIUrl":"10.1016/j.asoc.2026.114599","url":null,"abstract":"<div><div>Spectral convolutional neural networks using Fast Fourier Transform (FFT) often suffer from high computational complexity and memory demands due to complex valued operations and the need for inverse transforms limiting their deployment on resource constrained devices. This paper presents a novel end-to-end spectral convolutional neural network (SpCNN) architecture that operates entirely in the Discrete Cosine Transform (DCT) domain, eliminating the need for inverse transformations and complex arithmetic. Leveraging the DCT’s real valued representation and superior energy compaction, the proposed design significantly reduces computational workload and memory usage while preserving classification accuracy. Key innovations include the removal of IFFT layers, a frequency domain adaptive activation function (FReLU), and a DCT optimized spectral pooling mechanism, each tailored for deployment in low power, resource constrained environments. Experimental evaluations on MNIST and a 94-class ASCII dataset demonstrate the model’s efficiency: LeNet5-DCT achieves a 37.96% FLOPs reduction, 18.45% lower memory usage, and 96.56% test accuracy, while VGG7-DCT achieves a 33.95% FLOPs reduction, 14.32% lower memory usage, and 90.62% test accuracy. The architecture also shows strong robustness to quantization, confirming its suitability for edge AI applications and low energy inference. This work provides a scalable, hardware efficient spectral learning framework, paving the way for future hybrid spectral models optimized for embedded environments.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114599"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039850","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Multi-view feature representation learning: Integrating medical team information to enhance online physician recommendation 多视图特征表示学习:整合医疗团队信息,增强在线医生推荐
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-02-03 DOI: 10.1016/j.asoc.2026.114776
Yongbo Ni, Donghui Yang
Recommending suitable physicians to patients in online healthcare communities remains a challenging decision-making task. While existing studies leverage physicians’ professional features (clinical expertise) or service features (patient evaluations) for recommendations, single-view feature analysis and limited physician-patient interaction data often lead to insufficient feature representation, which further results in relatively poor recommendation performance. The emergence of online medical teams (OMTs) provides patients with a novel team perspective and data source for understanding physician features. Therefore, this study aims to leverage the OMTs information to construct a multi-view feature representation learning framework to enhance the physician feature representation and improve the recommendation performance. We first integrate the OMTs information with online reviews, disease descriptions, and physician profiles to build the multi-source and multi-type medical data inputs. Based on these inputs, we leverage BERT, TextCNN, semantic analysis, and graph neural network technologies to construct the multi-view feature representation learning framework from the views of professional features, service features, and team features. Then, these multi-view features are fused through a multilayer perceptron to predict the selection probabilities. Experimental results demonstrate the multi-view features based method incorporating OMTs information shows an average improvement of 7.5 % in HR@5, 5.3 % in NDCG@5, and 7.8 % in MRR compared to other classical methods. Surprisingly, the physician service capabilities involved in the team features have a greater impact on the recommendation performance than the physician professional capabilities involved.
向在线医疗保健社区的患者推荐合适的医生仍然是一项具有挑战性的决策任务。现有研究利用医生的专业特征(临床专业知识)或服务特征(患者评价)进行推荐,但单一视角的特征分析和有限的医患互动数据往往导致特征表达不足,从而导致推荐效果相对较差。在线医疗团队(omt)的出现为患者提供了一个新的团队视角和数据来源,以了解医生的特点。因此,本研究旨在利用omt信息构建一个多视图特征表征学习框架,以增强医生特征表征,提高推荐性能。我们首先将omt信息与在线评论、疾病描述和医生档案集成在一起,以构建多源、多类型的医疗数据输入。基于这些输入,我们利用BERT、TextCNN、语义分析和图神经网络技术,从专业特征、服务特征和团队特征的角度构建了多视图特征表示学习框架。然后,通过多层感知器融合这些多视图特征来预测选择概率。实验结果表明,与其他经典方法相比,结合OMTs信息的基于多视图特征的方法在HR@5、NDCG@5和7.8 %的MRR上分别平均提高了7.5 %、5.3 %和7.8 %。令人惊讶的是,团队特征中涉及的医生服务能力比涉及的医生专业能力对推荐绩效的影响更大。
{"title":"Multi-view feature representation learning: Integrating medical team information to enhance online physician recommendation","authors":"Yongbo Ni,&nbsp;Donghui Yang","doi":"10.1016/j.asoc.2026.114776","DOIUrl":"10.1016/j.asoc.2026.114776","url":null,"abstract":"<div><div>Recommending suitable physicians to patients in online healthcare communities remains a challenging decision-making task. While existing studies leverage physicians’ professional features (clinical expertise) or service features (patient evaluations) for recommendations, single-view feature analysis and limited physician-patient interaction data often lead to insufficient feature representation, which further results in relatively poor recommendation performance. The emergence of online medical teams (OMTs) provides patients with a novel team perspective and data source for understanding physician features. Therefore, this study aims to leverage the OMTs information to construct a multi-view feature representation learning framework to enhance the physician feature representation and improve the recommendation performance. We first integrate the OMTs information with online reviews, disease descriptions, and physician profiles to build the multi-source and multi-type medical data inputs. Based on these inputs, we leverage BERT, TextCNN, semantic analysis, and graph neural network technologies to construct the multi-view feature representation learning framework from the views of professional features, service features, and team features. Then, these multi-view features are fused through a multilayer perceptron to predict the selection probabilities. Experimental results demonstrate the multi-view features based method incorporating OMTs information shows an average improvement of 7.5 % in HR@5, 5.3 % in NDCG@5, and 7.8 % in MRR compared to other classical methods. Surprisingly, the physician service capabilities involved in the team features have a greater impact on the recommendation performance than the physician professional capabilities involved.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114776"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191514","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Evolving explainable neural architectures: Dynamic behavioral-novelty weighting in multi-objective neuroevolution for multimodal spam detection 进化的可解释神经结构:多目标神经进化中多模态垃圾邮件检测的动态行为新颖性加权
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-02-05 DOI: 10.1016/j.asoc.2026.114790
Adnane Filali , El Arbi Abdellaoui Alaoui , Mostafa Merras
Spam detection is effectively a moving target: as spammers shift their strategies across images, voice calls, and text, static detection systems struggle to keep pace. Artificial neural architectures, though often accurate, frequently fail to strike a workable balance between performance, generalization, and behavioral novelty in such adversarial environments. In this paper, we introduce and analyze two multi-objective variants of NeuroEvolution of Augmenting Topologies: a static weighting approach (NEAT-stMO) and a dynamic version (NEAT-dyMO) that adaptively reallocates objective priorities during the evolutionary process. Both methods simultaneously optimize for accuracy, the generalization gap, and novelty while evolving the topology and hyperparameters of artificial neural networks. We benchmark these evolutionary strategies against a broad spectrum of baselines, ranging from Pareto-based and single-objective NEAT variants to conventional classifiers, MLP, SVM, and automated deep learning frameworks, including Vision Transformers, Auto-Keras, and Hyperband. Our evaluation spans three distinct modalities—image, voice, and imbalanced SMS data—using five-fold cross-validation to rigorously assess accuracy, F1-scores, AUC, and computational efficiency—time and RAM. Importantly, we integrate SHapley Additive exPlanations to deconstruct the “black box” of how hyperparameters drive performance. Empirical results place NEAT-dyMO consistently among the top performers; notably, it offers significant efficiency advantages over heavy Transformer models while surpassing static baselines. SHAP analysis further sheds light on the mechanism, exposing that NEAT-dyMO’s dynamic weighting successfully navigates the trade-off between exploration—driven by larger architectures and dropout—and exploitation—favored by small batch sizes. These findings, therefore, suggest that dynamically evolving ANNs provide a robust, adaptive, and interpretable alternative for real-world multimodal spam defense.
垃圾邮件检测实际上是一个移动的目标:随着垃圾邮件发送者在图像、语音呼叫和文本之间改变策略,静态检测系统难以跟上步伐。人工神经结构虽然通常是准确的,但在这种敌对的环境中,往往无法在性能、泛化和行为新颖性之间取得可行的平衡。本文介绍并分析了增强拓扑神经进化的两种多目标变体:静态加权方法(NEAT-stMO)和动态版本(NEAT-dyMO),该方法在进化过程中自适应地重新分配目标优先级。这两种方法在进化人工神经网络的拓扑和超参数的同时,对精度、泛化差距和新颖性进行了优化。我们对这些进化策略进行了广泛的基准测试,从基于帕累托的单目标NEAT变体到传统分类器、MLP、SVM和自动深度学习框架,包括视觉变形器、Auto-Keras和Hyperband。我们的评估跨越了三种不同的模式——图像、语音和不平衡的短信数据——使用五倍交叉验证来严格评估准确性、f1分数、AUC和计算效率——时间和RAM。重要的是,我们整合了SHapley加性解释来解构超参数如何驱动性能的“黑匣子”。实证结果表明,NEAT-dyMO始终名列前茅;值得注意的是,与重型Transformer模型相比,它提供了显著的效率优势,同时超越了静态基线。SHAP分析进一步揭示了机制,揭示了NEAT-dyMO的动态加权成功地在探索(由大型体系结构驱动)和放弃(由小批量规模支持)之间进行权衡。因此,这些发现表明,动态进化的人工神经网络为现实世界的多模态垃圾邮件防御提供了一种鲁棒、自适应和可解释的替代方案。
{"title":"Evolving explainable neural architectures: Dynamic behavioral-novelty weighting in multi-objective neuroevolution for multimodal spam detection","authors":"Adnane Filali ,&nbsp;El Arbi Abdellaoui Alaoui ,&nbsp;Mostafa Merras","doi":"10.1016/j.asoc.2026.114790","DOIUrl":"10.1016/j.asoc.2026.114790","url":null,"abstract":"<div><div>Spam detection is effectively a moving target: as spammers shift their strategies across images, voice calls, and text, static detection systems struggle to keep pace. Artificial neural architectures, though often accurate, frequently fail to strike a workable balance between performance, generalization, and behavioral novelty in such adversarial environments. In this paper, we introduce and analyze two multi-objective variants of NeuroEvolution of Augmenting Topologies: a static weighting approach (NEAT-stMO) and a dynamic version (NEAT-dyMO) that adaptively reallocates objective priorities during the evolutionary process. Both methods simultaneously optimize for accuracy, the generalization gap, and novelty while evolving the topology and hyperparameters of artificial neural networks. We benchmark these evolutionary strategies against a broad spectrum of baselines, ranging from Pareto-based and single-objective NEAT variants to conventional classifiers, MLP, SVM, and automated deep learning frameworks, including Vision Transformers, Auto-Keras, and Hyperband. Our evaluation spans three distinct modalities—image, voice, and imbalanced SMS data—using five-fold cross-validation to rigorously assess accuracy, F1-scores, AUC, and computational efficiency—time and RAM. Importantly, we integrate SHapley Additive exPlanations to deconstruct the “black box” of how hyperparameters drive performance. Empirical results place NEAT-dyMO consistently among the top performers; notably, it offers significant efficiency advantages over heavy Transformer models while surpassing static baselines. SHAP analysis further sheds light on the mechanism, exposing that NEAT-dyMO’s dynamic weighting successfully navigates the trade-off between exploration—driven by larger architectures and dropout—and exploitation—favored by small batch sizes. These findings, therefore, suggest that dynamically evolving ANNs provide a robust, adaptive, and interpretable alternative for real-world multimodal spam defense.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114790"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191779","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
DMEL: A novel dual-modal ensemble learning architecture for multi-step runoff prediction DMEL:一种用于多步径流预测的新型双模态集成学习体系结构
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-02-06 DOI: 10.1016/j.asoc.2026.114792
Wen-chuan Wang , Miao Gu, Zong Li, Yi-yang Wang, Yan-wei Zhao
Accurate multi-step runoff prediction is crucial for effective water resource management and flood disaster reduction. Despite significant advancements in deep learning for hydrological predictions, identifying the modal characteristics of runoff sequences and extending the effective lead time of prediction models remain challenging. This paper introduces a novel Dual-Modal Ensemble Learning (DMEL) architecture that addresses these issues by integrating a Modal Recognition Strategy (MRS) with two deep learning models: Informer and Long Short-Term Memory (LSTM). The MRS technique first reduces the nonlinearity and non-stationarity of runoff data through feature extraction and fusion, lowering model computational complexity and mitigating cumulative errors from high-dimensional inputs. The optimized modal components are then categorized into high-frequency and low-frequency components, which are processed by the Informer and LSTM models, respectively, leveraging their unique capabilities for dual-modal ensemble learning. The final prediction is obtained by superimposing the results from all sequences. The two sites’ analysis results demonstrate the DMEL architecture’s effectiveness. Taking the Shuangpai Station as an example, in single-step prediction, the Kling-Gupta Efficiency (KGE) index of the DMEL model achieves relative improvements of 22.60 %, 28.32 %, 25.34 %, 43.90 %, 45.06 %, and 42.30 % compared to the Informer model, Transformer model, LSTM model, Recurrent Neural Network (RNN) model, TimesNet model, and PatchTST model, respectively. When the prediction step increases to the seventh step, the KGE index of the DMEL model still shows relative increases of 37.50 %, 105.71 %, 89.84 %, 146.04 %, 211.82 %, and 178.06 % compared to the Informer model, Transformer model, LSTM model, RNN model, TimesNet model, and PatchTST model, respectively. This indicates that the DMEL architecture can accurately capture critical information in daily runoff sequences, extending the model’s effective prediction period and providing a robust scientific basis for water resource management and flood disaster reduction.
准确的多步径流预测对于有效的水资源管理和减少洪水灾害至关重要。尽管深度学习在水文预测方面取得了重大进展,但识别径流序列的模态特征和延长预测模型的有效提前期仍然具有挑战性。本文介绍了一种新的双模态集成学习(DMEL)架构,该架构通过将模态识别策略(MRS)与两个深度学习模型:信息者和长短期记忆(LSTM)集成来解决这些问题。MRS技术首先通过特征提取和融合降低径流数据的非线性和非平稳性,降低模型计算复杂度,减轻高维输入的累积误差。然后将优化的模态分量分类为高频和低频分量,分别由Informer和LSTM模型处理,利用其独特的双模态集成学习功能。将所有序列的结果叠加得到最终的预测结果。两个站点的分析结果验证了DMEL体系结构的有效性。以双排站为例,在单步预测中,DMEL模型的KGE指数相对于Informer模型、Transformer模型、LSTM模型、RNN模型、TimesNet模型和PatchTST模型分别提高了22.60 %、28.32 %、25.34 %、43.90 %、45.06 %和42.30 %。当预测步长增加到第7步时,与Informer模型、Transformer模型、LSTM模型、RNN模型、TimesNet模型和PatchTST模型相比,DMEL模型的KGE指数仍然相对增加了37.50 %、105.71 %、89.84 %、146.04 %、211.82 %和178.06 %。这表明DMEL架构可以准确捕获日径流序列中的关键信息,延长了模型的有效预测周期,为水资源管理和洪水减灾提供了有力的科学依据。
{"title":"DMEL: A novel dual-modal ensemble learning architecture for multi-step runoff prediction","authors":"Wen-chuan Wang ,&nbsp;Miao Gu,&nbsp;Zong Li,&nbsp;Yi-yang Wang,&nbsp;Yan-wei Zhao","doi":"10.1016/j.asoc.2026.114792","DOIUrl":"10.1016/j.asoc.2026.114792","url":null,"abstract":"<div><div>Accurate multi-step runoff prediction is crucial for effective water resource management and flood disaster reduction. Despite significant advancements in deep learning for hydrological predictions, identifying the modal characteristics of runoff sequences and extending the effective lead time of prediction models remain challenging. This paper introduces a novel Dual-Modal Ensemble Learning (DMEL) architecture that addresses these issues by integrating a Modal Recognition Strategy (MRS) with two deep learning models: Informer and Long Short-Term Memory (LSTM). The MRS technique first reduces the nonlinearity and non-stationarity of runoff data through feature extraction and fusion, lowering model computational complexity and mitigating cumulative errors from high-dimensional inputs. The optimized modal components are then categorized into high-frequency and low-frequency components, which are processed by the Informer and LSTM models, respectively, leveraging their unique capabilities for dual-modal ensemble learning. The final prediction is obtained by superimposing the results from all sequences. The two sites’ analysis results demonstrate the DMEL architecture’s effectiveness. Taking the Shuangpai Station as an example, in single-step prediction, the Kling-Gupta Efficiency (KGE) index of the DMEL model achieves relative improvements of 22.60 %, 28.32 %, 25.34 %, 43.90 %, 45.06 %, and 42.30 % compared to the Informer model, Transformer model, LSTM model, Recurrent Neural Network (RNN) model, TimesNet model, and PatchTST model, respectively. When the prediction step increases to the seventh step, the KGE index of the DMEL model still shows relative increases of 37.50 %, 105.71 %, 89.84 %, 146.04 %, 211.82 %, and 178.06 % compared to the Informer model, Transformer model, LSTM model, RNN model, TimesNet model, and PatchTST model, respectively. This indicates that the DMEL architecture can accurately capture critical information in daily runoff sequences, extending the model’s effective prediction period and providing a robust scientific basis for water resource management and flood disaster reduction.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114792"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191782","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
MSCNRNet: Non-dense residual network with multi-skip connection for image restoration MSCNRNet:用于图像恢复的非密集残差网络,具有多跳点连接
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-02-05 DOI: 10.1016/j.asoc.2026.114742
Yun Chen , Huanhuan Luo , Mengnan Shi , Xiaowen Ma , Hongtao Li , Qiang Yao
In recent years, residual dense network(RDN)-based models have become increasingly popular in image restoration (IR) due to their enhanced feature propagation and ability to resist overfitting. However, these models still have limitations: 1) Their convolutional blocks rely on dense connections, leading to redundant local feature capture and repetitive features. 2) Each layer in the block is a single convolutional layer, limiting representational power and local feature extraction. 3) Basic connections between blocks hinder effective global feature fusion, resulting in reconstructed images lacking details. To address the limitations of RDN-based models, this paper proposes a novel IR model Non-dense Residual Network with Multi-skip Connection for Image Restoration (MSCNRNet). It replaces traditional dense connections in RDNs with multi-skip connections to form multi-skip connection non-dense residual blocks (MSCNRBs), reducing redundant features. Each MSCNRB comprises two multi-skip connection convolutional layers (MSCNRConv) to enhance local feature extraction, with a feature enhancement layer (FEL) added after each block for better global feature fusion. Based on extensive experiments across 12 synthetic and real-world datasets, our MSCNRNet model outperforms current state-of-the-art (SOTA) models in image restoration (IR) tasks by approximately 0.1% to 5% in the peak signal-to-noise ratio (PSNR) and around 2% in the structural similarity index measure (SSIM). This work provides an efficient solution to the limitations of existing RDN-based models, offering a more effective approach for high-quality IR tasks.
近年来,基于残差密集网络(RDN)的模型由于其增强的特征传播能力和抗过拟合能力,在图像恢复(IR)中得到越来越广泛的应用。然而,这些模型仍然存在局限性:1)它们的卷积块依赖于密集的连接,导致冗余的局部特征捕获和重复的特征。2)块中的每一层都是一个单独的卷积层,限制了表征能力和局部特征提取。3)块之间的基本连接阻碍了有效的全局特征融合,导致重构图像缺乏细节。针对基于rdn模型的局限性,提出了一种新的图像恢复非密集残差网络(MSCNRNet)模型。它用多跳连接代替rdn中传统的密集连接,形成多跳连接非密集残差块(MSCNRBs),减少冗余特征。每个MSCNRB包括两个多跳连接卷积层(MSCNRConv)以增强局部特征提取,每个块后增加一个特征增强层(FEL)以更好地进行全局特征融合。基于对12个合成数据集和真实世界数据集的广泛实验,我们的MSCNRNet模型在图像恢复(IR)任务中优于当前最先进的(SOTA)模型,在峰值信噪比(PSNR)方面优于约0.1%至5%,在结构相似性指数测量(SSIM)方面优于约2%。这项工作为现有基于rdn的模型的局限性提供了有效的解决方案,为高质量的红外任务提供了更有效的方法。
{"title":"MSCNRNet: Non-dense residual network with multi-skip connection for image restoration","authors":"Yun Chen ,&nbsp;Huanhuan Luo ,&nbsp;Mengnan Shi ,&nbsp;Xiaowen Ma ,&nbsp;Hongtao Li ,&nbsp;Qiang Yao","doi":"10.1016/j.asoc.2026.114742","DOIUrl":"10.1016/j.asoc.2026.114742","url":null,"abstract":"<div><div>In recent years, residual dense network(RDN)-based models have become increasingly popular in image restoration (IR) due to their enhanced feature propagation and ability to resist overfitting. However, these models still have limitations: 1) Their convolutional blocks rely on dense connections, leading to redundant local feature capture and repetitive features. 2) Each layer in the block is a single convolutional layer, limiting representational power and local feature extraction. 3) Basic connections between blocks hinder effective global feature fusion, resulting in reconstructed images lacking details. To address the limitations of RDN-based models, this paper proposes a novel IR model Non-dense Residual Network with Multi-skip Connection for Image Restoration (MSCNRNet). It replaces traditional dense connections in RDNs with multi-skip connections to form multi-skip connection non-dense residual blocks (MSCNRBs), reducing redundant features. Each MSCNRB comprises two multi-skip connection convolutional layers (MSCNRConv) to enhance local feature extraction, with a feature enhancement layer (FEL) added after each block for better global feature fusion. Based on extensive experiments across 12 synthetic and real-world datasets, our MSCNRNet model outperforms current state-of-the-art (SOTA) models in image restoration (IR) tasks by approximately 0.1% to 5% in the peak signal-to-noise ratio (PSNR) and around 2% in the structural similarity index measure (SSIM). This work provides an efficient solution to the limitations of existing RDN-based models, offering a more effective approach for high-quality IR tasks.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114742"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190764","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Robustness and invariance of hybrid metaheuristics under objective function transformations 目标函数变换下混合元启发式算法的鲁棒性和不变性
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-02-09 DOI: 10.1016/j.asoc.2026.114799
Grzegorz Sroka , Sławomir T. Wierzchoń
This paper evaluates the robustness and structural invariance of hybrid population-based metaheuristics under various objective space transformations. A light-weight plug-and-play hybridization operator is applied to nineteen state-of-the-art algorithms – including differential evolution, particle swarm optimization, and recent bio-inspired methods – without modifying their internal logic. Benchmarking on the CEC-2017 suite across four dimensions (10, 30, 50, 100) is performed under five transformation types: baseline, translation, scaling, rotation, and constant shift. Statistical comparisons based on Wilcoxon and Friedman tests, Bayesian dominance analysis, and convergence trajectory profiling consistently show that differential-based hybrids (e.g., hybrid improved multi-operator differential evolution, hybrid success-history adaptive differential evolution, hybrid double mutational salp swarm algorithm) maintain high accuracy, stability, and invariance across all tested deformations. In contrast, classical algorithms – especially particle swarm optimization- and Harris hawks optimization - based variants –exhibit significant performance degradation under non-separable or distorted landscapes. The findings confirm the superiority of adaptive, structurally resilient hybrids for real-world optimization tasks subject to domain-specific transformations. However, the evidence is based on deterministic, continuous, bound-constrained single-objective CEC-2017 functions. The experiments use a fixed budget and five predefined transformations. Therefore, the conclusions should be interpreted within this controlled setting and may not directly generalize to nonlinear constrained, noisy, dynamic, or multi-objective problems without dedicated protocols. To partially bridge to practice, we also include a constrained helical compression spring design case study evaluated with the same methodology.
本文评估了基于混合种群的元启发式算法在各种目标空间变换下的鲁棒性和结构不变性。轻量级即插即用的杂交算子应用于19种最先进的算法,包括差分进化,粒子群优化和最近的生物启发方法,而无需修改其内部逻辑。CEC-2017套件在四个维度(10,30,50,100)上的基准测试在五种转换类型下执行:基线、转换、缩放、旋转和恒定移动。基于Wilcoxon和Friedman测试、贝叶斯优势分析和收敛轨迹分析的统计比较一致地表明,基于微分的混合算法(例如,混合改进的多算子微分进化、混合成功历史自适应微分进化、混合双突变salp群算法)在所有测试变形中都保持了较高的准确性、稳定性和不变性。相比之下,经典算法——尤其是粒子群优化——和基于哈里斯鹰优化的变体——在不可分离或扭曲的景观下表现出显著的性能下降。研究结果证实了自适应的、结构弹性的混合动力车在实际领域特定转换下的优化任务中的优越性。然而,证据是基于确定性的,连续的,有界约束的单目标CEC-2017函数。实验使用固定的预算和五个预定义的转换。因此,结论应在这种受控环境下解释,而不能直接推广到非线性约束、噪声、动态或无专用协议的多目标问题。为了部分地与实践相结合,我们还包括一个用相同方法评估的受限螺旋压缩弹簧设计案例研究。
{"title":"Robustness and invariance of hybrid metaheuristics under objective function transformations","authors":"Grzegorz Sroka ,&nbsp;Sławomir T. Wierzchoń","doi":"10.1016/j.asoc.2026.114799","DOIUrl":"10.1016/j.asoc.2026.114799","url":null,"abstract":"<div><div>This paper evaluates the robustness and structural invariance of hybrid population-based metaheuristics under various objective space transformations. A light-weight plug-and-play hybridization operator is applied to nineteen state-of-the-art algorithms – including differential evolution, particle swarm optimization, and recent bio-inspired methods – without modifying their internal logic. Benchmarking on the CEC-2017 suite across four dimensions (10, 30, 50, 100) is performed under five transformation types: baseline, translation, scaling, rotation, and constant shift. Statistical comparisons based on Wilcoxon and Friedman tests, Bayesian dominance analysis, and convergence trajectory profiling consistently show that differential-based hybrids (e.g., hybrid improved multi-operator differential evolution, hybrid success-history adaptive differential evolution, hybrid double mutational salp swarm algorithm) maintain high accuracy, stability, and invariance across all tested deformations. In contrast, classical algorithms – especially particle swarm optimization- and Harris hawks optimization - based variants –exhibit significant performance degradation under non-separable or distorted landscapes. The findings confirm the superiority of adaptive, structurally resilient hybrids for real-world optimization tasks subject to domain-specific transformations. However, the evidence is based on deterministic, continuous, bound-constrained single-objective CEC-2017 functions. The experiments use a fixed budget and five predefined transformations. Therefore, the conclusions should be interpreted within this controlled setting and may not directly generalize to nonlinear constrained, noisy, dynamic, or multi-objective problems without dedicated protocols. To partially bridge to practice, we also include a constrained helical compression spring design case study evaluated with the same methodology.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114799"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191879","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Dynamic weight matching TEB local path planning improved by PSO for car-like robots 基于粒子群算法的仿车机器人动态权匹配TEB局部路径规划
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-02-04 DOI: 10.1016/j.asoc.2026.114783
Yunlong Wang , Annan Tang , Wenfeng Li , Shaoke Wan , Rongcan Qiu , Xiaohu Li
Local path planning for car-like robots often relies on the Timed Elastic Band (TEB) algorithm, yet conventional implementations suffer from two critical limitations: insufficient trajectory smoothness resulting in jerky motion and mechanical wear, and high sensitivity to weight parameters requiring laborious manual tuning that fails to generalize across diverse scenarios. To address these challenges, this paper proposes a novel framework that enhances TEB through explicit smoothness constraints while eliminating manual tuning via Particle Swarm Optimization (PSO). We first augment the TEB objective function with higher-order smoothness and jerk constraints to fundamentally improve trajectory quality and dynamic feasibility. Second, we introduce a systematic offline-online framework where PSO discovers optimal weight configurations across 150 representative training scenarios, each characterized by a three-dimensional feature vector capturing narrowness, turning complexity, and heading deviation. The discovered scenario-weight mappings are stored in a knowledge base enabling real-time weight retrieval through efficient similarity matching during online operation. Comprehensive evaluation across three progressively challenging test scenarios with 90 independent runs demonstrates statistically significant improvements: PSO-TEB achieves 34–38% execution time reduction, 39–53% acceleration smoothness improvement, and 100% success rate in dynamic environments compared to 0–20% for baseline methods. Real-world experiments on a car-like robotic platform validate practical viability with sub-millisecond online weight matching overhead. By shifting computational burden offline and enabling scenario-adaptive parameter adjustment, our approach enhances robustness and practical applicability of TEB for car-like robots in complex real-world conditions.
类车机器人的局部路径规划通常依赖于定时弹性带(TEB)算法,但传统的实现存在两个关键限制:轨迹平滑度不足,导致运动不稳定和机械磨损,以及对权重参数的高灵敏度,需要费力的手动调整,无法推广到不同的场景。为了解决这些挑战,本文提出了一个新的框架,通过明确的平滑约束来增强TEB,同时通过粒子群优化(PSO)消除手动调整。首先,我们用高阶平滑和抖动约束来增强TEB目标函数,从根本上改善轨迹质量和动态可行性。其次,我们引入了一个系统的离线-在线框架,其中PSO在150个代表性训练场景中发现最佳权值配置,每个训练场景都有一个三维特征向量,可以捕获窄度、转弯复杂性和航向偏差。发现的场景权重映射存储在知识库中,通过在线操作期间有效的相似度匹配实现实时权重检索。通过90次独立运行的三个逐步具有挑战性的测试场景的综合评估显示了统计学上显著的改进:PSO-TEB的执行时间减少了34-38%,加速平稳性提高了39-53%,在动态环境中的成功率为100%,而基线方法的成功率为0-20%。在类似汽车的机器人平台上进行的实际实验验证了亚毫秒在线重量匹配开销的实际可行性。通过将计算负担转移到离线状态并启用场景自适应参数调整,我们的方法增强了TEB在复杂现实条件下对类车机器人的鲁棒性和实用性。
{"title":"Dynamic weight matching TEB local path planning improved by PSO for car-like robots","authors":"Yunlong Wang ,&nbsp;Annan Tang ,&nbsp;Wenfeng Li ,&nbsp;Shaoke Wan ,&nbsp;Rongcan Qiu ,&nbsp;Xiaohu Li","doi":"10.1016/j.asoc.2026.114783","DOIUrl":"10.1016/j.asoc.2026.114783","url":null,"abstract":"<div><div>Local path planning for car-like robots often relies on the Timed Elastic Band (TEB) algorithm, yet conventional implementations suffer from two critical limitations: insufficient trajectory smoothness resulting in jerky motion and mechanical wear, and high sensitivity to weight parameters requiring laborious manual tuning that fails to generalize across diverse scenarios. To address these challenges, this paper proposes a novel framework that enhances TEB through explicit smoothness constraints while eliminating manual tuning via Particle Swarm Optimization (PSO). We first augment the TEB objective function with higher-order smoothness and jerk constraints to fundamentally improve trajectory quality and dynamic feasibility. Second, we introduce a systematic offline-online framework where PSO discovers optimal weight configurations across 150 representative training scenarios, each characterized by a three-dimensional feature vector capturing narrowness, turning complexity, and heading deviation. The discovered scenario-weight mappings are stored in a knowledge base enabling real-time weight retrieval through efficient similarity matching during online operation. Comprehensive evaluation across three progressively challenging test scenarios with 90 independent runs demonstrates statistically significant improvements: PSO-TEB achieves 34–38% execution time reduction, 39–53% acceleration smoothness improvement, and 100% success rate in dynamic environments compared to 0–20% for baseline methods. Real-world experiments on a car-like robotic platform validate practical viability with sub-millisecond online weight matching overhead. By shifting computational burden offline and enabling scenario-adaptive parameter adjustment, our approach enhances robustness and practical applicability of TEB for car-like robots in complex real-world conditions.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114783"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191820","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
3D-NoduleNet: A comprehensive framework for benign and malignant pulmonary nodule classification 3D-NoduleNet:肺良恶性结节分类的综合框架
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-01-13 DOI: 10.1016/j.asoc.2026.114645
Huihui Yu , Qun Dai , Yanfu Wu
Accurate pulmonary nodule classification is essential for lung cancer diagnosis but poses significant challenges due to class imbalance, high appearance heterogeneity, and subtle inter-class variations. To tackle these challenges, this paper proposes 3D-NoduleNet, a clinical-semantics-driven, unified framework that integrates detection and segmentation into a task-oriented pipeline for nodule classification. This framework begins with the Region Proposal and Segmentation Module (RPSM), which adopts a detection-guided segmentation strategy. The region proposals generated by the 3D Faster R-CNN are used to crop ROIs, which are subsequently segmented by the 3D BCDU-Net into aligned, morphologically consistent masks—explicitly instantiating key clinical semantic cues: texture and morphology. Next, the Multi-view Feature Extraction Module (MFEM) employs a dual-path architecture under a refine-then-fuse scheme. Within each path, an internal cascade refines view-specific semantic features through multi-scale contextual aggregation and sequential attention, whose output is fused to leverage complementary clinical cues. Finally, the Pulmonary Nodule Classification Module (PNCM) adopts a dual-supervision strategy: the binary cross-entropy loss provides primary instance-level supervision, while the auxiliary supervised contrastive loss regularizes the feature space by clustering same-class nodules and separating different-class ones—enhancing separability and improving benign-malignant classification. Through this coordinated semantic flow, the modules form a tightly integrated framework, systematically transforming clinical semantics from instantiation to discrimination. Experiments on LIDC-IDRI demonstrate the superior and well-balanced performance of 3D-NoduleNet over existing methods, with achievements of 92.50% sensitivity and 90.63% F1-score, alongside competitive accuracy and specificity. Ablation studies validate the effectiveness of key components and strategies, highlighting its robustness and clinical applicability.
准确的肺结节分类对肺癌的诊断至关重要,但由于类别不平衡、外观异质性高和微妙的类别间差异,存在重大挑战。为了应对这些挑战,本文提出了3D-NoduleNet,这是一个临床语义驱动的统一框架,将检测和分割集成到面向任务的结节分类管道中。该框架从区域建议和分割模块(RPSM)开始,RPSM采用检测引导的分割策略。3D Faster R-CNN生成的区域建议用于裁剪roi,随后由3D BCDU-Net分割成对齐的、形态一致的掩模,明确实例化关键的临床语义线索:纹理和形态。其次,多视图特征提取模块(MFEM)采用先细化后融合的双路径架构。在每个路径中,内部级联通过多尺度上下文聚合和顺序注意来细化特定于视图的语义特征,其输出被融合以利用互补的临床线索。最后,肺结节分类模块(Pulmonary结节Classification Module, PNCM)采用双监督策略:二值交叉熵损失提供主要的实例级监督,而辅助监督对比损失通过聚类同类结节和分离不同类别结节来正则化特征空间,增强可分性,改善良恶性分类。通过这种协调的语义流,各模块形成一个紧密集成的框架,系统地将临床语义从实例化转化为辨析。在LIDC-IDRI上的实验表明,3D-NoduleNet的性能优于现有的方法,灵敏度为92.50%,f1评分为90.63%,准确性和特异性具有竞争力。消融研究验证了关键成分和策略的有效性,突出了其稳健性和临床适用性。
{"title":"3D-NoduleNet: A comprehensive framework for benign and malignant pulmonary nodule classification","authors":"Huihui Yu ,&nbsp;Qun Dai ,&nbsp;Yanfu Wu","doi":"10.1016/j.asoc.2026.114645","DOIUrl":"10.1016/j.asoc.2026.114645","url":null,"abstract":"<div><div>Accurate pulmonary nodule classification is essential for lung cancer diagnosis but poses significant challenges due to class imbalance, high appearance heterogeneity, and subtle inter-class variations. To tackle these challenges, this paper proposes 3D-NoduleNet, a clinical-semantics-driven, unified framework that integrates detection and segmentation into a task-oriented pipeline for nodule classification. This framework begins with the Region Proposal and Segmentation Module (RPSM), which adopts a detection-guided segmentation strategy. The region proposals generated by the 3D Faster R-CNN are used to crop ROIs, which are subsequently segmented by the 3D BCDU-Net into aligned, morphologically consistent masks—explicitly instantiating key clinical semantic cues: texture and morphology. Next, the Multi-view Feature Extraction Module (MFEM) employs a dual-path architecture under a refine-then-fuse scheme. Within each path, an internal cascade refines view-specific semantic features through multi-scale contextual aggregation and sequential attention, whose output is fused to leverage complementary clinical cues. Finally, the Pulmonary Nodule Classification Module (PNCM) adopts a dual-supervision strategy: the binary cross-entropy loss provides primary instance-level supervision, while the auxiliary supervised contrastive loss regularizes the feature space by clustering same-class nodules and separating different-class ones—enhancing separability and improving benign-malignant classification. Through this coordinated semantic flow, the modules form a tightly integrated framework, systematically transforming clinical semantics from instantiation to discrimination. Experiments on LIDC-IDRI demonstrate the superior and well-balanced performance of 3D-NoduleNet over existing methods, with achievements of 92.50% sensitivity and 90.63% F1-score, alongside competitive accuracy and specificity. Ablation studies validate the effectiveness of key components and strategies, highlighting its robustness and clinical applicability.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114645"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039729","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
An adaptive multi-stage clustering-based differential evolution algorithm with dynamic niching for robust solving of high-dimensional nonlinear equation systems 基于动态小生境的多阶段聚类自适应差分进化算法用于高维非线性方程组的鲁棒求解
IF 6.6 1区 计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE Pub Date : 2026-04-01 Epub Date: 2026-01-18 DOI: 10.1016/j.asoc.2026.114666
Wulfran Fendzi Mbasso , Hassan M. Hussein Farh , Ambe Harrison , Abdullrahman A. Al-Shamma
Solving high-dimensional systems of Nonlinear Equation Systems (NES) requires search procedures that preserve diversity while converging reliably and, where applicable, discovering multiple roots. We propose the Adaptive Multi-stage Clustering Differential Evolution (AMCDE) algorithm, which couples fitness-aware clustering with entropy-triggered dynamic niching and feeds cluster-level convergence signals into mutation/crossover control in Differential Evolution (DE). We evaluate AMCDE on 110 NES benchmarks under a matched computation budget (population size of 100 and 10,000 function evaluations) and a uniform success threshold ‖F(x)‖₂ ≤ 10⁻⁸ (relaxed to 10⁻⁶ for flagged ill-conditioned subsets). Distinct-root discovery is quantified using Euclidean (ℓ₂) distance–based deduplication at 10⁻³ (optional pre-cluster at 10⁻⁴). We compare against recent NES-oriented competitors and adaptive DE baselines, including JADE (Adaptive Differential Evolution with Optional External Archive) and SHADE (Success-History Adaptive Differential Evolution), re-implemented on the same platform with grid-tuned settings for fairness. AMCDE attains the best median success rate and residuals, with higher distinct-root coverage at comparable runtimes; nonparametric Wilcoxon tests and Friedman tests with Holm correction confirm significance. A critical-difference diagram places AMCDE in the top clique. These findings indicate that coupling adaptive clustering with dynamic niching yields a robust, scalable NES solver that preserves exploration without sacrificing convergence efficiency. All comparative results are validated using non-parametric statistics (Wilcoxon signed-rank and Friedman tests with Holm post-hoc, α=0.05), with our method showing statistically significant gains over competing DE variants across the benchmark suite.
求解非线性方程系统(NES)的高维系统要求搜索程序在保持多样性的同时可靠地收敛,并在适用的情况下发现多个根。提出了一种自适应多阶段聚类差分进化(AMCDE)算法,该算法将适应度感知聚类与熵触发的动态小生境相结合,并将聚类级收敛信号输入差分进化(DE)的突变/交叉控制中。我们在匹配的计算预算(人口规模为100和10,000个功能评估)和统一的成功阈值‖F(x)‖₂≤ 10⁻⁸(对于标记的病态子集松弛到10⁻⁶)下,在110个NES基准上评估AMCDE。在10⁻³ (可选的10⁻⁴预聚类)使用基于欧氏(r₂)距离的重复数据删除来量化不同的根发现。我们比较了最近面向nes的竞争对手和自适应DE基线,包括JADE(带有可选外部存档的自适应差异进化)和SHADE(成功历史自适应差异进化),它们在同一平台上重新实现,具有网格调整的公平性设置。AMCDE获得了最好的中位数成功率和残差,在可比运行时具有更高的不同根覆盖率;非参数Wilcoxon检验和带有Holm校正的Friedman检验证实了显著性。一个临界差图将AMCDE置于最顶端的集团中。这些发现表明,将自适应聚类与动态小生境相结合,可以产生一个鲁棒的、可扩展的NES求解器,在不牺牲收敛效率的情况下保持探索。所有比较结果都使用非参数统计数据进行验证(Wilcoxon符号秩和Friedman检验与Holm事后检验,α=0.05),我们的方法显示在基准套件中竞争DE变体的统计显著收益。
{"title":"An adaptive multi-stage clustering-based differential evolution algorithm with dynamic niching for robust solving of high-dimensional nonlinear equation systems","authors":"Wulfran Fendzi Mbasso ,&nbsp;Hassan M. Hussein Farh ,&nbsp;Ambe Harrison ,&nbsp;Abdullrahman A. Al-Shamma","doi":"10.1016/j.asoc.2026.114666","DOIUrl":"10.1016/j.asoc.2026.114666","url":null,"abstract":"<div><div>Solving high-dimensional systems of Nonlinear Equation Systems (NES) requires search procedures that preserve diversity while converging reliably and, where applicable, discovering multiple roots. We propose the Adaptive Multi-stage Clustering Differential Evolution (AMCDE) algorithm, which couples fitness-aware clustering with entropy-triggered dynamic niching and feeds cluster-level convergence signals into mutation/crossover control in Differential Evolution (DE). We evaluate AMCDE on 110 NES benchmarks under a matched computation budget (population size of 100 and 10,000 function evaluations) and a uniform success threshold ‖F(x)‖₂ ≤ 10⁻⁸ (relaxed to 10⁻⁶ for flagged ill-conditioned subsets). Distinct-root discovery is quantified using Euclidean (ℓ₂) distance–based deduplication at 10⁻³ (optional pre-cluster at 10⁻⁴). We compare against recent NES-oriented competitors and adaptive DE baselines, including JADE (Adaptive Differential Evolution with Optional External Archive) and SHADE (Success-History Adaptive Differential Evolution), re-implemented on the same platform with grid-tuned settings for fairness. AMCDE attains the best median success rate and residuals, with higher distinct-root coverage at comparable runtimes; nonparametric Wilcoxon tests and Friedman tests with Holm correction confirm significance. A critical-difference diagram places AMCDE in the top clique. These findings indicate that coupling adaptive clustering with dynamic niching yields a robust, scalable NES solver that preserves exploration without sacrificing convergence efficiency. All comparative results are validated using non-parametric statistics (Wilcoxon signed-rank and Friedman tests with Holm post-hoc, α=0.05), with our method showing statistically significant gains over competing DE variants across the benchmark suite.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114666"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039725","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
期刊
Applied Soft Computing
全部 Acc. Chem. Res. ACS Applied Bio Materials ACS Appl. Electron. Mater. ACS Appl. Energy Mater. ACS Appl. Mater. Interfaces ACS Appl. Nano Mater. ACS Appl. Polym. Mater. ACS BIOMATER-SCI ENG ACS Catal. ACS Cent. Sci. ACS Chem. Biol. ACS Chemical Health & Safety ACS Chem. Neurosci. ACS Comb. Sci. ACS Earth Space Chem. ACS Energy Lett. ACS Infect. Dis. ACS Macro Lett. ACS Mater. Lett. ACS Med. Chem. Lett. ACS Nano ACS Omega ACS Photonics ACS Sens. ACS Sustainable Chem. Eng. ACS Synth. Biol. Anal. Chem. BIOCHEMISTRY-US Bioconjugate Chem. BIOMACROMOLECULES Chem. Res. Toxicol. Chem. Rev. Chem. Mater. CRYST GROWTH DES ENERG FUEL Environ. Sci. Technol. Environ. Sci. Technol. Lett. Eur. J. Inorg. Chem. IND ENG CHEM RES Inorg. Chem. J. Agric. Food. Chem. J. Chem. Eng. Data J. Chem. Educ. J. Chem. Inf. Model. J. Chem. Theory Comput. J. Med. Chem. J. Nat. Prod. J PROTEOME RES J. Am. Chem. Soc. LANGMUIR MACROMOLECULES Mol. Pharmaceutics Nano Lett. Org. Lett. ORG PROCESS RES DEV ORGANOMETALLICS J. Org. Chem. J. Phys. Chem. J. Phys. Chem. A J. Phys. Chem. B J. Phys. Chem. C J. Phys. Chem. Lett. Analyst Anal. Methods Biomater. Sci. Catal. Sci. Technol. Chem. Commun. Chem. Soc. Rev. CHEM EDUC RES PRACT CRYSTENGCOMM Dalton Trans. Energy Environ. Sci. ENVIRON SCI-NANO ENVIRON SCI-PROC IMP ENVIRON SCI-WAT RES Faraday Discuss. Food Funct. Green Chem. Inorg. Chem. Front. Integr. Biol. J. Anal. At. Spectrom. J. Mater. Chem. A J. Mater. Chem. B J. Mater. Chem. C Lab Chip Mater. Chem. Front. Mater. Horiz. MEDCHEMCOMM Metallomics Mol. Biosyst. Mol. Syst. Des. Eng. Nanoscale Nanoscale Horiz. Nat. Prod. Rep. New J. Chem. Org. Biomol. Chem. Org. Chem. Front. PHOTOCH PHOTOBIO SCI PCCP Polym. Chem.
﹀
mjeditor
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
0
微信
客服QQ
Book学术公众号 扫码关注我们
反馈
×
意见反馈
请填写您的意见或建议
请填写您的手机或邮箱
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
现在去查看 取消
×
提示
确定
Book学术官方微信
Book学术官方微信
Book学术文献互助
Book学术文献互助群
群 号:604180095
Book学术
文献互助 智能选刊 最新文献 互助须知 联系我们:info@booksci.cn
Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。
Copyright © 2023 Book学术 All rights reserved.
ghs 京公网安备 11010802042870号 京ICP备2023020795号-1