Biomedical Signal Processing and Control最新文献

{"title":"Convolutional attention-based adaptive separation network for EEG artefact removal","authors":"Jiaqiang Wang ,&nbsp;Jiaxin Xie ,&nbsp;Pengrui Li ,&nbsp;Shihong Liu ,&nbsp;Xinmin Ding ,&nbsp;Jun Chen ,&nbsp;Lutao Wang ,&nbsp;Dongrui Gao","doi":"10.1016/j.bspc.2025.108320","DOIUrl":"10.1016/j.bspc.2025.108320","url":null,"abstract":"<div><div>Electroencephalography (EEG) is an important electrical signal for recording physiological activities of the brain. Usually, its high temporal resolution can be susceptible to contamination by artefacts such as electrooculogram (EOG) and electromyogram (EMG), which affects the subsequent signal processing and analysis. Therefore, EEG artefact removal is particularly important and can lay the foundation for brain–computer interface (BCI) applications. However, most of the existing studies suffer from insufficient ability to extract global and local coarse-grained and fine-grained features interactively, insufficient effective constraints on latent features, and excessive computational complexity. To address these problems, this study designs a convolutional attention-based adaptive separation network (ASNet) for EEG artefact removal. Here, a separator is introduced based on the U-net architecture to constrain the potential features so as to adaptively separate the required discriminative features. In addition, this study improves on the disadvantages of the Transformer itself by designing a novel convolutional attention module to improve the coarse-grained and fine-grained feature interaction learning capability while reducing the number of parameters and computational effort. The experimental results show that ASNet achieves impressive artefact removal performance on fully synthetic datasets, semi-synthetic datasets, and real datasets, and has low computational complexity, which is a significant advantage over existing state-of-the-art methods. The codes are available at <span><span>https://github.com/qwertwjq/ASNet/tree/main</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108320"},"PeriodicalIF":4.9,"publicationDate":"2025-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144595820","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}

{"title":"Dual teacher–student network with feature calibration and dynamic consistency for semi-supervised medical image segmentation","authors":"Jiwen Zhou ,&nbsp;Fabien Pfaender ,&nbsp;Wanyu Liu","doi":"10.1016/j.bspc.2025.108284","DOIUrl":"10.1016/j.bspc.2025.108284","url":null,"abstract":"<div><div>Semi-supervised learning plays a crucial role in medical image segmentation, leveraging limited labeled data and abundant unlabeled data to reduce annotation demands and improve model performance. Nonetheless, existing methods struggle with pseudo-label quality, ineffective consistency regularization, and error accumulation, especially in scenarios involving complex structures and unclear boundaries. To address these challenges, we propose DTCC-Net, a dual teacher–student framework with three targeted innovations. First, we introduce a high-dimensional feature calibration (HDFC) module to enhance encoder feature alignment using orthogonal vector selection, improving pseudo-label precision. Second, a student consistency (SC) module applies entropy-based adaptive consistency loss to better regulate uncertain regions, overcoming limitations of global consistency constraints. Finally, an ensemble tree (ET) module performs structure-aware decoupling using a minimum spanning tree to suppress early-stage error propagation. These designs address critical limitations of existing methods by enhancing encoder-level feature calibration, decoder-level uncertainty-aware consistency, and spatial structure modeling, thereby improving both training stability and segmentation generalization. The LA, Pancreas-CT, and FLARE datasets, which represent typical medical image characteristics, are used to evaluate DTCC-Net. The results show that DTCC-Net achieves superior performance across all evaluation metrics, and ablation studies along with hyperparameter analyses further confirm the effectiveness of our methods.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108284"},"PeriodicalIF":4.9,"publicationDate":"2025-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144595944","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}

{"title":"Exploring functional near-infrared spectroscopy and transfer learning for discriminating major depressive disorder and generalized anxiety disorder","authors":"Hele Liu ,&nbsp;Jinzhou Zhu ,&nbsp;Jitao Zhong ,&nbsp;Shi Qiao ,&nbsp;Lu Zhang ,&nbsp;Jiangang Li ,&nbsp;Bin Hu ,&nbsp;Sujie Ma ,&nbsp;Hong Peng","doi":"10.1016/j.bspc.2025.108070","DOIUrl":"10.1016/j.bspc.2025.108070","url":null,"abstract":"<div><div>The comorbidity between Generalized Anxiety Disorder (GAD) and Major Depressive Disorder (MDD) is notably high, underscoring the need for nuanced diagnostic tools to differentiate between the two. Despite the rich clinical landscape, automated tests for discerning MDD from GAD remain underdeveloped, with existing disparities in brain function between the disorders yet to be fully elucidated. Functional Near-Infrared Spectroscopy (fNIRS), a promising neuroimaging modality, holds the potential to bridge this gap by supplementing clinical interviews and mental status examinations. This study engaged 113 participants, including 39 individuals clinically diagnosed with GAD, 35 healthy controls (HCs), and 39 individuals clinically diagnosed with MDD, collecting resting-state fNIRS data over three minutes from each subject. Measurements of brain oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) were captured using a 22-channel continuous wave fNIRS device to investigate resting state brain dynamics. In a departure from conventional machine learning methodologies, which often suffer from lower accuracy and the need for manual feature selection, this research introduces a pioneering transfer learning strategy. The raw fNIRS time series data were transformed into images via the Gramian Angular Summation Field method, facilitating the subsequent classification task through the application of three distinct neural networks: EfficientNet, MobileNet, and ResNet. Given the significant influence of resting-state data length on functional connectivity outcomes, a methodical approach utilizing a 30-second step size was employed to identify the optimal time scale for functional connectivity analysis. This approach leveraged the average accuracy across the three models as a guiding criterion, ultimately selecting data durations that maximized classification accuracy for functional connectivity exploration. Moreover, this study proposes a novel data processing technique, Temporal Compression by Averaging (TCA), designed to enhance data stability and conserve computational resources. Data processed through TCA demonstrated robust performance in classification tasks, with an optimal duration of 150 s identified for the resting-state data. This duration achieved an average accuracy rate of 90.42% across the networks, peaking at 93.82%. Functional connectivity analyses conducted on this optimized data subset revealed distinct connectivity patterns in patient groups relative to HCs. Notably, GAD patients exhibited increased functional connectivity strength in the MFG.R and ORBsup.R regions compared to MDD patients, highlighting the intricate neural underpinnings of these disorders. These insights affirm the potential of fNIRS as a valuable adjunct in the clinical diagnosis of comorbid MDD and GAD, showcasing the transformative impact of transfer learning methodologies in advancing fNIRS research.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108070"},"PeriodicalIF":4.9,"publicationDate":"2025-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144595505","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}

{"title":"GPDENet: Combining gradient-path design and dynamic enhancement strategy for medical image segmentation","authors":"Yakun Yang,&nbsp;Lin Li,&nbsp;Longhe Wang,&nbsp;Hongcheng Xue,&nbsp;Xupeng Kou","doi":"10.1016/j.bspc.2025.108313","DOIUrl":"10.1016/j.bspc.2025.108313","url":null,"abstract":"<div><div>The precise and automated segmentation of medical images is crucial in disease diagnosis and treatment planning. Recently, the transformer has emerged as a successful tool for capturing long-range dependencies, but migrating it to dense prediction tasks presents challenges due to the square complexity associated with input resolution. The low-level detailed features are essential for medical segmentation, as the boundaries of region-of-interest are prone to mis-segmentation. Inadequate local–global interaction of each scale is harmful for object to propagate cues. In this paper, we propose the GPDENet that adopts a canonical u-shape design for medical image segmentation. The encoder optimizes its compositions from the aspect of gradient path analysis, while integrating long-range context through a query-independent global vector with minimal overhead. Additionally, the decoder employs a progressively boosted tactic to adjust decoding features and applies the feature compression to reduce complexity from quadratic to approximate polynomial quantities. Finally, the light-weight full-scale skip connections are introduced to interlink the encoder with decoder, promoting feature reuse and alleviating semantic gaps. We evaluate our GPDENet on public computed tomography datasets and its performance consistently surpasses the majority of existing methods. Specifically, it outperforms ScaleFormer with an improvement from 0.7483 to 0.8321 mIoU on Kidney Parsing 2022, and it outperforms MTUNet in mDice from 0.7621 to 0.8468 on AbdomenCT-1K. Our approach mitigates the heavy reliance on pre-trained weights for transformers and overcomes the constraint of CNNs in capturing long-range relationships. Moreover, the cascade dynamic enhancement improves feature quality, generating complete and precise organ masks.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108313"},"PeriodicalIF":4.9,"publicationDate":"2025-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144595817","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}

{"title":"Multiple temporal scale network for remote PPG and heart rate estimation from facial video","authors":"Dao Q. Le ,&nbsp;Wen-Nung Lie ,&nbsp;Po-Han Huang ,&nbsp;Guan-Hao Fu ,&nbsp;Quynh Nguyen Quang Nhu","doi":"10.1016/j.bspc.2025.108309","DOIUrl":"10.1016/j.bspc.2025.108309","url":null,"abstract":"<div><div>This study presents a deep-learning-based multi-temporal-scale network that is named SlowFast TMB Net to estimate remote photoplethysmogram (rPPG) signals and heart rate values using a facial video sequence. The network consists of three 2D-convolutional streams: one spatial stream extracts spatial attention masks and two temporal streams capture the temporal dependency modelling using the same video input at different frame rates. The attention masks from the spatial stream are used for weighting information to enhance the feature maps for temporal streams. A fusion block is then used to combine the weighted feature maps for different temporal scales to create the final features to estimate the rPPG. This 3-stream (and fusion) architecture accurately determines the diverse temporal dependency characteristics. This study also proposes a sliding-window process for the test or inference stage for post-processing to predict the rPPG signals with redundancies. This procedure addresses the issues of environmental noise or user motion. The experimental results using four public datasets (PURE, MMSE-HR, UBFC-rPPG and MAHNOB-HCI) show that the proposed method gives results that are comparable to or better than those for state-of-the-art methods, which demonstrates the network’s learning capability. (<span><span>https://github.com/ccudsp520/SlowFast-TMB-Net-rPPG</span><svg><path></path></svg></span>)</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108309"},"PeriodicalIF":4.9,"publicationDate":"2025-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144595818","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}

{"title":"Hybrid feature fusion and randomized regression network for brain age prediction using magnetic resonance imaging: Analyzing age-related cortical structure variations in healthy adults","authors":"Raveendra Pilli ,&nbsp;Tripti Goel ,&nbsp;R. Murugan ,&nbsp;Alzheimer’s Disease Neuroimaging Initiative","doi":"10.1016/j.bspc.2025.108255","DOIUrl":"10.1016/j.bspc.2025.108255","url":null,"abstract":"<div><div>The disparity between real age and predicted age computed using machine learning/deep learning and magnetic resonance imaging (MRI), known as the brain age gap (BAG), serves as an indicator for predicting neurocognitive disorders. While effective at extracting local features, Convolutional neural networks (CNNs) often struggle to capture global dependencies essential for accurate brain age estimation. In contrast, the Vision Transformer (ViT) model captures global features via self-attention, modeling relationships across the entire image. In this study, a brain age estimation framework is developed using T1-weighted structural MRI data from three publicly available neuroimaging databases. 1,070 healthy control (HC) subjects are used for the age prediction model training and testing. Brain age prediction is performed using deep features extracted directly from whole-brain MRI scans via ViT and ResNet-50, without relying on handcrafted morphometry features. These features are concatenated and input into an <span><math><mrow><mi>l</mi><mn>1</mn></mrow></math></span>-regularized random vector functional link (RVFL) regression network, encouraging model sparsity and reducing overfitting by minimizing irrelevant feature contributions. The proposed model achieved state-of-the-art performance, with a mean absolute error (MAE) of 2.34 years and 3.21 years of root mean square error (RMSE). To evaluate BAG as a potential biomarker for assessing brain health, independent testing is conducted on 240 subjects from the ADNI database, including 100 with mild cognitive impairment (MCI), and 140 with Alzheimer’s disease (AD). The findings revealed that individuals affected by Alzheimer’s exhibit a higher rate of error outcomes, suggesting accelerated brain aging. Additionally, age-related cortical changes such as increased surface variability and right-hemisphere cortical thinning are observed in older individuals compared to younger subjects.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108255"},"PeriodicalIF":4.9,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144595769","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}

{"title":"Enhancing human activity recognition with TB-ConvAtt: A multi-dimensional attention framework","authors":"Hongmei Yang ,&nbsp;Yan Wang ,&nbsp;Ruixiang Hu ,&nbsp;Yingrui Geng ,&nbsp;Aihui Wang ,&nbsp;Xiaohu Zhou ,&nbsp;Hongnian Yu ,&nbsp;Qiangsong Zhao","doi":"10.1016/j.bspc.2025.108314","DOIUrl":"10.1016/j.bspc.2025.108314","url":null,"abstract":"<div><div>The growing prevalence of wearable technology in healthcare highlights the essential need for accurate and efficient human activity recognition (HAR) using wearable sensor data. In clinical settings, HAR plays a pivotal role in patient monitoring, rehabilitation, and personalized healthcare management. This study introduces TB-ConvAtt, a lightweight and multi-dimensional framework that integrates Convolutional Neural Networks (CNNs) with specialized attention mechanisms to effectively balance the extraction of independent temporal, spatial, and spatio-temporal features from wearable multi-sensor data. TB-ConvAtt consists of three distinct branches: the Temporal Attention Dimension (TAD), the Spatial Attention Dimension (SAD), and the Spatio-temporal Attention Dimension (STAD). The performance of TB-ConvAtt is thoroughly evaluated on four public HAR datasets (UNIMIB-SHAR, OPPORTUNITY, PAMAP2, MHEALTH). Comparative studies and detailed ablation experiments demonstrate that TB-ConvAtt achieves state-of-the-art performance while maintaining a lightweight design, enabling efficient deployment in resource-constrained environments.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108314"},"PeriodicalIF":4.9,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144595492","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}

{"title":"SES-Glass: A Smart Eye Sensor based Wearable Retinal Glass for AMD disease classification via bi-directional CNN","authors":"Aisha Banu W. ,&nbsp;Sandhya M. ,&nbsp;Arputha Rathina X. ,&nbsp;Leninisha Shanmugam","doi":"10.1016/j.bspc.2025.108206","DOIUrl":"10.1016/j.bspc.2025.108206","url":null,"abstract":"<div><div>Age-related macular degeneration (AMD) is an advanced eye ailment mainly affecting central vision, potentially leading to vision impairment and blindness. However, recognition of AMD at an early stage is critical to slowing its progression. In this paper, a novel SES-Glass: A Smart Eye Sensor-based Wearable Retinal Glass has been designed for AMD disease classification, which predicts AMD at its early stage and prevents vision loss. The proposed SES-Glass uses smart glasses equipped with a Raspberry Pi, an infrared sensor, and microphones to measure eye gaze. The Chebyshev filter is applied to the collected acoustic signals to enhance them by eliminating the noisy distortions. Deep learning based Bidirectional Convolutional Recurrent Neural Networks (BCRNN) is used to analyse the noise-free acoustic signals to classify the four severity stages of AMD namely early, intermediate, advanced, and normal. Moreover, the Healthcare professionals are informed of predicted results and patients receive notifications about their condition. For the experimental analysis, the specific metrics like precision, accuracy, recall, specificity and F1 score are used to evaluate the efficiency of the proposed SES-Glass for classifying the AMD stages. The proposed SES-Glass attains a maximum accuracy of 98.45% based on the gathered acoustic eye signal for the efficient classification of AMD stages. According to the findings, the accuracy of the proposed SES-Glass increased by 0.45<!--> <!-->0.90%, and 3.05% compared to the existing 13-Layer DCNNAMDNet23, ODL, and 13-Layer DCNN respectively.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108206"},"PeriodicalIF":4.9,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144588270","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}

{"title":"Simultaneous pharmacometric control optimization for advanced integrated multi-drug target infusions","authors":"Pablo Martinez-Vazquez ,&nbsp;Ana Abad-Torrent","doi":"10.1016/j.bspc.2025.108271","DOIUrl":"10.1016/j.bspc.2025.108271","url":null,"abstract":"<div><div>Accurately controlling drug delivery is crucial for safe anesthesia. Target-controlled infusion (TCI) systems use pharmacokinetic and pharmacodynamic (PK/PD) models to administer intravenous agents to reach target concentrations. However, TCI’s operation is restricted to a single-drug, not accounting for drug interactions. Dose–response interaction (DRI) models quantify such interactions by representing shared effects as a function of agents’ concentrations. We introduce a new administering methodology for multi-drug infusions, interaction target-controlled infusion (iTCI), that combines the PK/PD models of the co-administered drugs and their interactions into a single optimal non-linear dynamic control problem with terminal constraints. The capabilities of the iTCI are shown in different clinical scenarios using different PK/PD models such as Schnider and Eleveld for propofol and Minto for remifentanil. Incorporating DRI and PK/PD models allows novel administration procedures. These show that: (1) iTCI requires lower administered volumes than TCI to simultaneously reach the same target concentrations. (2) It offers optimal interdependent administrations that address not only concentration targets but also effect targets. (3) iTCI comes with additional constraints on the administration, including controlled titrations along iso-effect conditions (isoboles) or (4) directly limiting plasma concentration levels. (5) Unlike TCI, iTCI can include in the control strategy different exerted effects (multiple ke0s) per drug, particularly relevant for opioids. The iTCI is a new versatile open-loop multi-drug infusion paradigm allowing safer and more flexible optimized concurrent administration profiles than traditional TCI solutions. It opens up new possibilities for simultaneous multi-anesthetics controlled administrations where synergies and effects are taken into consideration.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108271"},"PeriodicalIF":4.9,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144588273","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}

{"title":"Robust semantic learning for precise medical image segmentation","authors":"Snehashis Chakraborty ,&nbsp;Komal Kumar ,&nbsp;Ankan Deria ,&nbsp;Dwarikanath Mahapatra ,&nbsp;Behzad Bozorgtabar ,&nbsp;Sudipta Roy","doi":"10.1016/j.bspc.2025.108251","DOIUrl":"10.1016/j.bspc.2025.108251","url":null,"abstract":"<div><div>Precisely localizing anomalies in medical images remains a significant challenge due to their heterogeneous nature across modalities and organs. While initial efforts excelled in identifying prominent anomalies, detecting minute target lesions posed significant limitations. These minute anomalies are particularly elusive and demand advanced detection techniques. Additionally, many existing models demand high computational resources, limiting their practicality in real-world clinical settings. In this study, we present REUnet, a novel Unet based architecture designed to address these obstacles by providing precise segmentation while also exhibiting strong generalization across diverse modalities and organs. The core advantage of REUnet resides in its resilient encoding pathway, constructed upon a module called dynamic mobile inverted bottleneck convolution. This module introduces a gating signal that significantly enhances semantic information, enabling the model to focus on specific regions of interest. The encoding pathway of REUnet is also linked strategically with the decoder to ensure efficient processing of these robust features which facilitates better communication between the two. Furthermore, the use of depth-wise separable convolution and dropout layers further makes REUnet computationally efficient for clinical use. Extensive experiments conducted on five publicly available datasets, including DUKE, BRATS2020, KiTS2023, INBreast, and FracAtlas, demonstrate REUnet’s strong generalization capabilities and superior performance, establishing a new state-of-the-art in medical image segmentation. The source code is available at GitHub link: <span><span>https://github.com/labsroy007/RobustSemanticLearning</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"110 ","pages":"Article 108251"},"PeriodicalIF":4.9,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144595771","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}