Biomedical Signal Processing and Control最新文献

{"title":"Deep learning-based computer-aided diagnostic system for lumbar degenerative diseases classification using MRI","authors":"Yueyao Chen ,&nbsp;Qiangtai Huang ,&nbsp;Chu Zhang ,&nbsp;Junfeng Li ,&nbsp;Wensheng Huang ,&nbsp;Peiyin Luo ,&nbsp;Qiuyi Chen ,&nbsp;Ruirui Qi ,&nbsp;Yuxuan Wan ,&nbsp;Bingsheng Huang ,&nbsp;Zhenhua Gao ,&nbsp;Xiaofeng Lin ,&nbsp;Songxiong Wu ,&nbsp;Xianfen Diao","doi":"10.1016/j.bspc.2025.108002","DOIUrl":"10.1016/j.bspc.2025.108002","url":null,"abstract":"<div><div>Lumbar degenerative diseases (LDDs) are prevalent orthopedic conditions worldwide, presenting significant diagnostic challenges due to high patient volumes in many healthcare institutions. To address this challenge, we developed Lumbar CAD, a deep learning-based computer-aided diagnostic (CAD) system for classifying LDDs using dual-view MRI. A lumbar disc localization and region extraction method based on nnUNetv2 was implemented to extract disc-level images from patient-level MRI data. Subsequently, sagittal lumbar disc images were analyzed using a binary classification model to predict degenerative discs. Axial images were then processed using a multi-label classification model to classify seven distinct types of lumbar disc lesions. To enhance system development efficiency, labels for lumbar degenerative diseases were obtained from clinical diagnostic reports instead of requiring re-annotation by radiologists. Data from two medical centers were collected for training and validation of the Lumbar CAD system. The results showed that the Lumbar CAD system achieved patient-level disc localization success rates of 96.7 % and 98.7 %, and disc-level multi-label classification accuracies of 0.890 and 0.879 on the internal and external test sets, respectively. Additionally, on the internal test set, the system per-patient LDD diagnoses approximately 120 times faster than experienced radiologists (1.09 ± 0.01 s vs. 132.41 ± 63.22 s per case). These findings highlight the potential of Lumbar CAD to assist clinicians in the accurate and efficient diagnosis of LDDs.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 108002"},"PeriodicalIF":4.9,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144068155","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":"Automated volumetric assessment of hepatocellular carcinomas using multi-phase-fused dual-attention network","authors":"Nuo Tong ,&nbsp;Shuiping Gou ,&nbsp;Shanshan Xu ,&nbsp;Yanyan Zhou ,&nbsp;Yunbao Cao ,&nbsp;Jing Yi ,&nbsp;Shixin Chen ,&nbsp;Hu Zhang ,&nbsp;Junde Liu ,&nbsp;Jing Yu ,&nbsp;Zhonghua Luo","doi":"10.1016/j.bspc.2025.108025","DOIUrl":"10.1016/j.bspc.2025.108025","url":null,"abstract":"<div><h3>Objectives</h3><div>Assessing patients’ responses to Transarterial Chemoembolization (TACE) treatment by analyzing the volume changes of Hepatocellular Carcinomas (HCC) is critical for treatment planning and prognosis. To obtain the precise cancer volumes in real time and assess its responses efficiently, an automated segmentation network is proposed for HCCs and extensively evaluated in this study.</div></div><div><h3>Methods</h3><div>To tackle the low image contrast in the single phase CT images, a network based on the fusion of the multi-phase images and dual-attention mechanism (MFDA-Net) is developed. Despite the U-Net-based backbone network, the Multi-Phase Feature Fusion (MPFF) and Dual-Attention Mechanism (DAM) are proposed to fuse the semantic features from different phases and enhance the spatial and channel-wise discriminative features, respectively. To automate the assessment of the HCC patients’ responses, the correlations between the cancer volumes produced by the segmentation results and the manual delineations were deeply analyzed using diverse metrics.</div></div><div><h3>Results</h3><div>Six metrics were employed for the quantitative evaluation of the proposed MFDA-Net on a multi-phase abdominal CT dataset. Furthermore, Response Evaluation Criteria in Solid Tumors (RECIST) and volumetric RECIST (vRECIST) were utilized to analyze the cancer volumes. Experimental results demonstrated that the proposed MFDA-Net outperforms the other twelve comparison algorithms significantly. Meanwhile, the segmentation results have strong statistical correlations with the manual delineations on both of the RECIST and vRECIST.</div></div><div><h3>Conclusions</h3><div>Extensive experiments validated that the proposed MFDA-Net is an optional method compared with the labor-intensive manual delineations to assess the HCC patients’ responses after the treatments. The source code is available at <span><span>https://github.com/mqy-git111/MFDA-Net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 108025"},"PeriodicalIF":4.9,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144068158","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":"Bayes factors for sequential auditory brainstem response detection","authors":"M.A. Chesnaye ,&nbsp;D.M. Simpson ,&nbsp;J. Schlittenlacher ,&nbsp;S. Laugesen ,&nbsp;S.L. Bell","doi":"10.1016/j.bspc.2025.107937","DOIUrl":"10.1016/j.bspc.2025.107937","url":null,"abstract":"<div><h3>Objective</h3><div>When determining the presence or absence of an Auditory Brainstem Response (ABR), clinicians often visually inspect the accruing data over time, i.e., a sequential test is adopted. The current work presents and evaluates Bayes Factors (BFs) as an objective sequential test for assisting clinicians with this task.</div></div><div><h3>Method</h3><div>Test specificity and sensitivity were optimised in simulated data and evaluated in subject-recorded data, including no-stimulus recordings (17 adults) and chirp-evoked ABR recordings (31 adults, 9 with hearing loss). The BF approach was compared with an existing sequential test, called the Convolutional Group Sequential Test (CGST).</div></div><div><h3>Results</h3><div>In simulations, BFs reduced mean test times by 60–70 % relative to the CGST while maintaining equal sensitivity and specificity. Similar reductions were observed in subject-recorded EEG background activity (∼70 %) and in chirp-evoked ABRs (0–60 %, depending on stimulus levels). For BFs, test time is tied to noise levels in the data, which allows test sensitivity to be controlled even when noise levels are high. The drawback is that the FPR is also tied to test time, and results show small variations (&lt;0.01) in FPRs depending on noise levels. In contrast, test time for the CGST is fixed, giving an improved control over the FPR, but a reduced control over test sensitivity.</div></div><div><h3>Significance</h3><div>BFs demonstrated high sensitivity and reduced mean test times relative to the CGST. It also provides regular feedback with no maximum test time specified, making it well-suited at assisting clinicians with different levels of expertise and feedback preferences.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"108 ","pages":"Article 107937"},"PeriodicalIF":4.9,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143946477","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":"Adaptive auxiliary diffusion for multi-modal brain tumor segmentation with random missing modalities","authors":"Boyue Zhao ,&nbsp;Yi Zhang ,&nbsp;Meng Zhao ,&nbsp;Guoxia Xu ,&nbsp;Congcong Wang","doi":"10.1016/j.bspc.2025.108015","DOIUrl":"10.1016/j.bspc.2025.108015","url":null,"abstract":"<div><div>Brain tumor segmentation methods based on multi-modal MRI perform significantly well when data is complete. In clinical settings, the absence of modalities due to artifacts and equipment problems often renders these methods ineffective. Current research attempts to train a universal model to adapt to 15 different random combinations of missing modalities. However, due to the random and complex nature of missing modality combinations across different cases, a single model faces challenges in dynamically adjusting its processing strategy to accommodate specific missing modality scenarios, ultimately leading to diminished segmentation accuracy. In this work, we introduce an end-to-end Incomplete Multi-modal Diffusion Brain Tumor Segmentation (IMD-TumorSeg) framework, which is designed to handle various scenarios with missing modalities. Specifically, the model incorporates independent generative modules for each modality and introduces an adaptive conditional integration mechanism to dynamically adjust the weight fusion between missing and available modalities. In addition, an attention-driven diffusion strategy is proposed to facilitate collaborative learning between the diffusion process and the segmentation network. Furthermore, by integrating an image estimator, the framework evaluates the similarity between generated and real images in real-time, optimizing the generation process and ensuring both visual and semantic consistency of the generated images. Extensive experimental results on the BraTS 2018 and BraTS 2020 datasets demonstrate that IMD-TumorSeg exhibits superior performance and effectiveness in handling missing modalities compared to state-of-the-art methods.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 108015"},"PeriodicalIF":4.9,"publicationDate":"2025-05-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143943225","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 healthcare system for patient data with HEO-PRAFT consensus mechanism using SPS authentication based EPRSS encryption","authors":"Ashish Kumar Chakraverti ,&nbsp;Kavita Saini ,&nbsp;Gaurav Raj ,&nbsp;Shwetav Sharad ,&nbsp;Murari Kumar Singh","doi":"10.1016/j.bspc.2025.108050","DOIUrl":"10.1016/j.bspc.2025.108050","url":null,"abstract":"<div><div>Owing to the advantages of immutability, blockchain has been proposed as a potential solution to facilitate the transmission of personal health data while maintaining anonymity and security. With its decentralized, unchangeable, and transparent characteristics, blockchain technology presents a viable way for addressing these issues. In proposed sophisticated EP (Elliptic curve-Proxy-Re-encryption) based RSS encryption (Ring Signature Scheme) system and Human Evolutionary Optimization (HEO) based effective Practical Byzantine Fault Tolerance (PBFT) enabled RAFT (HEO-PRAFT) is developed for secured EMR sharing and storing in blockchain. The suggested approach is to obtain patient information from standard hospitals’ Electronic Medical Records (EMRs). Subsequently, the gathered data are verified by the use of a structural preserving scheme (SPS), and the verified data are encrypted through the EP-RSS scheme. A public–private key pair for the user is generated in advance EPRSS scheme. The encrypted EMR data is stored in the cloud. Then HEO-PRAFT based consensus process is used to hash the encrypted data and add it to the blockchain. The healthcare data is safely transferred to the blockchain for effective and secure patient data storage based on this enhanced PRAFT. The method uses EPRSS-based access control to ensure privacy protection and offer customized access control. The obtained experimental outcomes, including transaction delay, block creation time, encryption time, decryption times, jitter, and throughput, are 4.51 ms, 3.71 ms, 4.56 ms, 4.85 ms, 4.09 ms, and 539mb/s. Ultimately, EPRSS and HEO-PRAFT based encryption and consensus methods enhance the security of the healthcare system.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 108050"},"PeriodicalIF":4.9,"publicationDate":"2025-05-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143943224","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":"Computer Aided Diagnosis of lung cancer using deep learning techniques","authors":"Agees Kumar C. ,&nbsp;Merry Ida A.","doi":"10.1016/j.bspc.2025.107994","DOIUrl":"10.1016/j.bspc.2025.107994","url":null,"abstract":"<div><div>The most critical stages in improving patient outcomes are diagnosis as well as earlier detection of lung cancer. Technology is crucial in detecting cancer. Based on their findings, many researchers have presented various techniques. Several Computer-Aided Diagnostic (CAD) approaches and systems have been presented and developed in recent years to address this difficulty using computer technology. The Lung Image Database Consortium-Image Database Resource Initiative (LIDC-IDRI) dataset was utilized to evaluate the effectiveness of the suggested method. To start, pre-processing the raw images lowers the noise in the final dataset. Pre-processing is done with a Median Filter (MF). The segmented data is examined using the Fuzzy Weighted Local Information C Means (FWLICM) Segmentation technique in order to identify the lung cancer’s affected region. The feature is extracted using a model called the scale invariant feature transform (SIFT). The proposed study uses the CNN-LSTM architecture to categorize brain tumors as “BENIGN” or “MALIGNANT”. The proposed CNN-LSTM model improves precision and accurately diagnoses lung cancer. The proposed research developed a hybrid deep CNN-LSTM network with the accuracy of 99.49% to significantly support earlier lung diagnosis of cancer among individuals.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 107994"},"PeriodicalIF":4.9,"publicationDate":"2025-05-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143946697","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":"Physiologically inspired spatiotemporal adaptive multimodal fusion model for blood glucose prediction","authors":"Yifan Yin ,&nbsp;Zhihao Zhao ,&nbsp;Jingtong Na ,&nbsp;Chundong Xue ,&nbsp;Kairong Qin","doi":"10.1016/j.bspc.2025.107998","DOIUrl":"10.1016/j.bspc.2025.107998","url":null,"abstract":"<div><div>Blood glucose (BG) prediction is critical to achieving efficient and stable BG management for diabetic patients. Recently, data-driven BG prediction methods based on deep learning have achieved excellent results. However, most current methods overlook the fact that BG is regulated by a variety of physiological factors, and it is difficult and physiologically unrealistic to reflect future BG changes solely from historical BG data. Therefore, we develop a <strong>S</strong>patio<strong>T</strong>emporal <strong>A</strong>daptive (STA) model that deeply integrates physiologically BG metabolism principles with deep learning for multimodal BG prediction. Temporally, an independent modality reversal attention mechanism is proposed by combining the time delay differences in the impact of different exogenous modalities on BG in metabolic dynamics, ensuring the physiological independence of different exogenous modalities to avoid feature entanglement between them and independently extracting the temporal correlations between them and BG. Spatially, the framework implements hierarchical processing by distinguishing between direct and indirect physiological effects of exogenous modalities on BG levels. A multi-level multimodal fusion and modality candidate mechanism are proposed to handle the difference between the influence of inter-type modality and intra-type modality. This approach builds a physiologically inspired deep learning model from both spatial and temporal perspectives, providing a new avenue for subsequent research. Our method has been well validated on real-world datasets, achieving an average RMSE of <span><math><mrow><mn>18</mn><mo>.</mo><mn>43</mn><mo>±</mo><mn>2</mn><mo>.</mo><mn>40</mn></mrow></math></span> for 30 min prediction horizons and <span><math><mrow><mn>29</mn><mo>.</mo><mn>73</mn><mo>±</mo><mn>3</mn><mo>.</mo><mn>88</mn></mrow></math></span> for 60 min prediction horizons.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 107998"},"PeriodicalIF":4.9,"publicationDate":"2025-05-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143943223","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":"EFDFNet: A multimodal deep fusion network based on feature disentanglement for attention state classification","authors":"Haiqi Xu ,&nbsp;Qingshan She ,&nbsp;Ming Meng ,&nbsp;Yunyuan Gao ,&nbsp;Yingchun Zhang","doi":"10.1016/j.bspc.2025.108042","DOIUrl":"10.1016/j.bspc.2025.108042","url":null,"abstract":"<div><div>The classification of attention states utilizing both electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) is pivotal in understanding human cognitive functions. While multimodal algorithms have been explored within brain-computer interface (BCI) research, the integration of modal features often falls short of efficacy. Moreover, comprehensive multimodal classification studies employing deep learning techniques for attention state classification are limited. This paper proposes a novel EEG-fNIRS multimodal deep fusion framework (EFDFNet), which employs fNIRS features to enhance EEG feature disentanglement and uses a deep fusion strategy for effective multimodal feature integration. Additionally, we have developed EMCNet, an attention state classification network for the EEG modality, which combines Mamba and Transformer to optimize the extraction of EEG features. We evaluated our method on two attention state classification datasets and one motor imagery dataset, i.e., mental arithmetic (MA), word generation (WG) and motor imagery (MI). The results show that EMCNet achieved classification accuracies of 86.11%, 79.47% and 75.77% on the MA, WG and MI datasets using only the EEG modality. With multimodal fusion, EFDFNet improved these results to 87.31%, 80.90% and 85.61%, respectively, highlighting the benefits of multimodal fusion. Both EMCNet and EFDFNet deliver state-of-the-art performance and are expected to set new baselines for EEG-fNIRS multimodal fusion.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 108042"},"PeriodicalIF":4.9,"publicationDate":"2025-05-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143943221","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":"EchoSAM: SAM adaption for unified 2D echocardiography segmentation and ejection fraction calculation","authors":"Xue Li ,&nbsp;Qian Hu ,&nbsp;Xiangbo Lin ,&nbsp;Yushi Li ,&nbsp;Yu Dong ,&nbsp;Tong Lin","doi":"10.1016/j.bspc.2025.108000","DOIUrl":"10.1016/j.bspc.2025.108000","url":null,"abstract":"<div><div>Automatic segmentation of echocardiography and the calculation of clinical data play a crucial role in the assessment of cardiac function. The left ventricular ejection fraction (LVEF) is a key indicator of the heart’s systolic performance. In this study, we present EchoSAM, a unified framework designed for integrated structures segmentation and LVEF calculation based on the Segment Anything Model (SAM). SAM exhibits strong and precise zero-shot segmentation skills in natural images. Nevertheless, because of the domain difference, it cannot be applied to echocardiography. In order to mitigate the effects of blurred boundaries, low contrast, and high noise, we enhance the Image Encoder to acquire more informative features. Meanwhile, according to the LVEF calculation of Simpson method, we design a points localization module, leveraging the combination of image features and Mask Decoder output to obtain precise points locations. Our EchoSAM model not only enables an accurate LVEF calculation in a fully automatic way, but also allows checking the segmentation quality of cardiac structures to ensure clear and reliable clinical analysis. We rigorously evaluated our approach on three datasets: CAMUS, EchoNet-Dynamic, and EchoDUT. The results demonstrate that EchoSAM has achieved Dice of 94.03% for the left ventricle (LV), 88.48% for the myocardium (MYO), superior to other state-of-the-art methods. Additionally, LVEF yielded a mean absolute error (MAE) of 6.39% and a root mean square error (RMSE) of 8.56%.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 108000"},"PeriodicalIF":4.9,"publicationDate":"2025-05-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143936309","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":"Multi-domain fusion network: A novel approach to OCTA image segmentation in diabetic retinopathy","authors":"Guangmei Jia ,&nbsp;Fei Ma ,&nbsp;Sien Li ,&nbsp;Zhaohui Zhang ,&nbsp;Hongjuan Liu ,&nbsp;Yanfei Guo ,&nbsp;Jing Meng","doi":"10.1016/j.bspc.2025.107945","DOIUrl":"10.1016/j.bspc.2025.107945","url":null,"abstract":"<div><div>Retinal lesions signify a cascade of pathological alterations disrupting normal retinal function, which makes their automatic segmentation pivotal for the timely detection of eye diseases. Optical coherence tomography angiography (OCTA) non-invasively visualizes the 3D vascular structure of the retina by analyzing blood flow signals at different depths. However, early microvascular lesions, such as microaneurysms, show minimal changes in blood flow, rendering them subtle and easily missed in OCTA images. Furthermore, the unique OCTA imaging process introduces inherent stripe noise. Existing deep learning-based segmentation algorithms mainly rely on spatial domain information from a single network, making it challenging to accurately capture such subtle changes. To address this problem, we propose a Multi-Domain Fusion Network (MFNet) that captures both spatial and frequency domain features from OCTA images for retinal lesion segmentation. It is the first time to design a novel frequency domain encoder by fusing multi-level Discrete Wavelet Transform (DWT), capturing multi-scale texture features while reducing noise. Moreover, we design a Domain Fusion Module (DFM) that employs a multi-level fusion strategy and gating mechanism to fully integrate spatial and frequency features, addressing the shortcomings of simple concatenation or addition in existing methods. Experimental results show that MFNet outperforms current methods on multiple datasets. For example, on the Diabetic Retinopathy Analysis Challenge 2022 (DRAC2022) dataset, MFNet achieved dice coefficients of 54.48% and 75.36% for neovascularization, intraretinal microvascular abnormalities, and non-perfusion areas, with intersection over union (IoU) values of 65.02%, 37.44%, and 60.47%, respectively. Our code is available at <span><span>https://github.com/GM-JIa/MFNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"109 ","pages":"Article 107945"},"PeriodicalIF":4.9,"publicationDate":"2025-05-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143937264","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}