Biomedical Signal Processing and Control最新文献

{"title":"xU-NetFullSharp: The Novel Deep Learning Architecture for Chest X-ray Bone Shadow Suppression","authors":"","doi":"10.1016/j.bspc.2024.106983","DOIUrl":"10.1016/j.bspc.2024.106983","url":null,"abstract":"<div><h3>Background and objectives</h3><div>Chest X-ray image (CXR) is vital for screening, preventing, and monitoring various lung diseases. In particular, the early detection of lung cancer can significantly improve patients’ chances of survival and quality of life. Unfortunately, approximately 82–95 % of missed pulmonary nodules are estimated to be obscured by rib shadows, making them difficult to recognize. This study addresses this problem by considering the rib shadows in CXRs as noise that can be reduced using deep learning. The result of the proposed model is a CXR with improved clarity for easier and more accurate analysis by radiologists or computer algorithms.</div></div><div><h3>Methods</h3><div>An automated deep learning-based model for bone shadow suppression from frontal CXRs, called xU-NetFullSharp, was proposed. This network is inspired by the most modern U-NetSharp architecture and was modified using different approaches to preserve as many details as possible and accurately suppress bone shadows. For comparison, recent state-of-the-art models were implemented and trained. JSRT, VinDr-CXR, and Gusarev DES datasets were utilized for the experiments, where the JSRT dataset was extensively augmented.</div></div><div><h3>Results</h3><div>The performance of the proposed xU-NetFullSharp was analyzed using statistical measures and compared with that of other architectures. The proposed model significantly outperformed the others, reaching the best values of the most used metrics (0.9846 SSIM; 0.9870 MS-SSIM). It also achieves a correlation of 96.31 % and an intersection of 10.0285 between the predicted and ground truth histograms, together with the smallest value of the Bhattacharyya distance. The obtained results were validated by experts from the University Hospital Olomouc with positive feedback, thus achieving the best objective and subjective results. The proposed method has the potential to be implemented in hospital environments.</div></div><div><h3>Conclusion</h3><div>A comprehensive comparison of the proposed architecture with state-of-the-art methods proves its efficiency in suppressing noise in CXRs and its ability to distinguish the signals of important tissues from noise components. This methodology can potentially improve the performance of the existing CXR processing methods. The source code is released in a GitHub repository that can be accessed from the following link: <span><span><u>https://github.com/xKev1n/xU-NetFullSharp</u></span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142445754","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":"Automatic semantic segmentation of breast cancer in DCE-MRI using DeepLabV3+ with modified ResNet50","authors":"","doi":"10.1016/j.bspc.2024.106691","DOIUrl":"10.1016/j.bspc.2024.106691","url":null,"abstract":"<div><div>Research on breast cancer segmentation is essential due to its high prevalence as the most common cancer in women and its occurrence in men as well. Breast cancer involves abnormal cell growth in the breast, highlighting the importance of advanced imaging. Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCE-MRI) is an effective technique for this purpose. Deep learning has significantly influenced medical imaging in recent years, especially in accurately segmenting tumors from MRI images. Two techniques have been proposed for breast tumor segmentation: Dilated ResNet50 (RN50D) and Parallel Layers Added ResNet50 (PLA-RN50). RN50D involves altering the dilation factor of the convolution layer within the residual block of ResNet50. PLA-RN50 entails the integration of parallel layers following the final residual block of the ResNet50 architecture. The modified architectures serve as the backbone for the DeepLabV3+ network. The DeepLabV3+ with RN50D or PLA-RN50D is a powerful and effective architecture that integrates deep feature extraction, multiscale spatial information, and precise segmentation to achieve high accuracy in lesion segmentation for breast DCE-MRI images. The proposed technique is tested on a QIN Breast DCE-MRI dataset comprising 233 images sourced from The Cancer Image Archive. The proposed method achieves a dice score of 0.92. The superior segmentation performance of DeepLabV3+ with PLA-RN50, as compared to its counterparts using ResNet18 and ResNet50, highlights the impactful modifications incorporated in PLA-RN50 for optimizing breast tumor segmentation.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142445728","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":"Deep vein thrombosis detection via combination of neural networks","authors":"","doi":"10.1016/j.bspc.2024.106972","DOIUrl":"10.1016/j.bspc.2024.106972","url":null,"abstract":"<div><div>Deep Vein Thrombosis (DVT) is the result of blood clots in the veins of the body especially in the legs. The most catastrophic importance of DVT is pulmonary embolism (PE), which is caused by a portion of the clot breaking off and entering the bloodstream and lungs. However, the clinical diagnosis of DVT is time consuming so if a computer aided system is available it will be really efficient. In this study, a novel Deep R-Belief network is proposed to identify DVT in Duplex Ultrasound (DUS) images. Initially, the input DUS images are denoised with the gaussian filter to remove the noisy distortions and Contrast Limited Adaptive Histogram Equalization (CLAHE) technique is applied to these images for improving the image quality. Then, the noise-free images are given as input to the fuzzy based threshold algorithm for segmenting the edges. The deep learning based RegNet is utilized for extracting the most relevant features from the segmented output images. After that, Deep belief Network (DBN) is applied for classifying the DUS images into coronary thrombosis, venous thrombo embolism and pulmonary embolism. The competence of the Deep R-Belief network was assessed by metrics like precision, specificity, accuracy, recall, and F1 score. From the experimental findings, the accuracy of the proposed Deep R-Belief network method is 98.63%. The efficiency of the Deep R-Belief network advances the overall accuracy value by 9.7%, 18.8% and 15.8% better than DL-CNN, SESARF and XGBoost respectively.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142441724","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":"Early stage brain tumor prediction using dilated and Attention-based ensemble learning with enhanced Artificial rabbit optimization for brain data","authors":"","doi":"10.1016/j.bspc.2024.107033","DOIUrl":"10.1016/j.bspc.2024.107033","url":null,"abstract":"<div><div>The integration of deep learning into brain data analysis has notably boosted the field of biomedical data analysis. In the context of intricate conditions like cancer, various data modalities can reveal distinct disease characteristics. Brain data has the potential to expose additional insights compared to using the data sources in isolation. Moreover, techniques are selected and prioritized based on the speed and accuracy of the data. Therefore, a new deep learning technique is assisted in predicting the brain tumor from the brain data to provide accurate prediction outcomes. The brain data required for predicting the brain tumor is garnered through various online sources. Then, the collected data are applied to the data preprocessing phase for cleaning the collected brain data and then applied to the data transformation method to improve the efficiency for providing better decision-making over prediction. The transformed data is then offered to the weighted feature selection process, where the weights of the features are optimized through the proposed Enhanced Artificial Rabbits Optimizer. The selection of weighted features is primarily adopted for solving the data dimensionality issues and these resultant features are given to the Dilated and Attention-based Ensemble Learning Network to provide the effective prediction outcome, where the deep learning structures like 1-Dimensional Convolutional Neural Networks, Bidirectional Long Short-Term Memory (BiLSTM), Deep Temporal Convolution Network are ensembled in the DAEL network. Finally, the prediction outcome attained from the proposed model is validated through the existing brain tumor prediction frameworks to ensure the efficacy of the implemented scheme.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142446866","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":"A DF-SSA analytical framework for revealing variations in multidimensional EEG features of epileptic seizures","authors":"","doi":"10.1016/j.bspc.2024.107073","DOIUrl":"10.1016/j.bspc.2024.107073","url":null,"abstract":"<div><div>Epilepsy, recognized as the most prevalent chronic neurological disorder globally, markedly impacts affected individuals’ lives. Epileptic states are effectively detected by electroencephalography (EEG), yet the potential of machine learning to epilepsy’s neural mechanisms remains underutilized. In this study, an analytical framework employing the Deep Forest (DF) classifier and Sparrow Search Algorithm (SSA) was introduced to extract an optimal subset of multidimensional EEG features from the CHB-MIT and JHMCHH datasets, aimed at revealing significant variations in epilepsy. Three pivotal epileptic states-interictal, preictal, and ictal-were identified, with Relative Power (RP), Sample Entropy (SE), and Mutual Information (MI) computed for each. The importance ranking and selection of features were facilitated by the DF-SSA framework, leading to the identification of an optimal subset that achieved notable classification accuracies of 98.38 ± 0.42 % and 99.09 ± 0.91 %, which represent increases of 0.56 % and 2.04 % over the baseline, respectively. Additionally, significant changes within the beta and gamma bands across the three states were revealed by analyzing variations in cerebral cortex activity, with SE showcasing consistent patterns and a marked elevation from the interictal to preictal, and finally to ictal periods. Surprisingly, SE more readily distinguished the three epileptic states than RP, due to its sensitivity to signal complexity changes. Additionally, a reorganization of functional connectivity across all brain regions was uncovered to be triggered by seizures. Through this innovative analytical framework’s employment, three key epileptic seizure states were identified, revealing significant variations in brain electrical features and offering insights into epilepsy’s complexity.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142446867","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":"Time–frequency domain machine learning for detection of epilepsy using wearable EEG sensor signals recorded during physical activities","authors":"","doi":"10.1016/j.bspc.2024.107041","DOIUrl":"10.1016/j.bspc.2024.107041","url":null,"abstract":"<div><div>Epilepsy is a neurological ailment in which there is a disturbance in the nerve cell activity of the brain, causing recurrent seizures. The electroencephalogram (EEG) signal is widely used as a diagnostic modality to detect epilepsy ailment. The automated detection of epilepsy using wearable EEG sensor data recorded during various physical activities is interesting for continuous monitoring of brain health. This paper proposes a time–frequency (TF) domain machine learning (ML) approach for the automated detection of epilepsy using wearable sensor-based EEG signals. The Gaussian window-based Stockwell transform (GWST) is employed to evaluate the TF matrix from the EEG signal. The features such as the <span><math><msub><mrow><mi>L</mi></mrow><mrow><mn>1</mn></mrow></msub></math></span>-norm and the Shannon entropy are extracted from the TF matrix of the EEG signal. The ML and deep learning (DL) models are employed to detect epilepsy using the TF domain features of EEG signals. The publicly available database containing wearable sensor-based EEG signals recorded from the subjects while performing different physical activities is used to evaluate the performance of the proposed approach. The results show that the random forest (RF) classifier coupled with GWST domain features of EEG signals has obtained an overall accuracy value of 90.74% for detecting epilepsy with hold-out validation using the EEG signals from different physical activity cases. For the 10-fold cross-validation (CV) case, the GWST domain features of EEG signal and multi-layer long short-term memory (LSTM) classifier have produced the average accuracy value of 74.44%. For jogging, running, and idle sitting activities, the GWST-based TF domain entropy features coupled with the multi-layer LSTM model have obtained accuracy values of 82.72%, 82.41%, and 87.30%, respectively. The proposed approach has achieved higher classification accuracy than existing methods to detect epilepsy using wearable sensor-based EEG signals using a 10-fold CV strategy. The suggested approach is compared with existing methods to classify seizure and seizure-free classes using resting-state EEG signals.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142441181","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":"X-Brain: Explainable recognition of brain tumors using robust deep attention CNN","authors":"","doi":"10.1016/j.bspc.2024.106988","DOIUrl":"10.1016/j.bspc.2024.106988","url":null,"abstract":"<div><div>Automated brain tumor recognition is crucial for swift diagnosis and treatment in healthcare, enhancing patient survival rates but manual recognition of tumor types is time-consuming and resource-intensive. Over the past few years, researchers have proposed various Deep Learning (DL) methods to automate the recognition process over the past years. However, these approaches often lack convincing accuracy and rely on datasets consisting of limited samples, raising concerns regarding real-world efficacy and reliability. Furthermore, the decisions made by black-box AI models in healthcare, where lives are at stake, require proper decision explainability. To address these issues, we propose a robust and explainable deep CNN-based method for effective recognition of brain tumor types, attaining state-of-the-art accuracies of 99.81%, 99.55%, and 99.30% on the training, validation, and test sets, respectively, surpassing both the previous works and baseline models. Moreover, we employed three Explainable AI techniques: Grad-CAM, Grad-CAM++, and Score-CAM for explainability analysis, contributing towards the development of trustworthy and reliable automation of healthcare diagnosis.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142441725","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":"Non-invasive prediction of atrial fibrillation recurrence by recurrence quantification analysis on the fibrillation cycle length","authors":"","doi":"10.1016/j.bspc.2024.107037","DOIUrl":"10.1016/j.bspc.2024.107037","url":null,"abstract":"<div><h3>Objective:</h3><div>The long-term success of atrial fibrillation (AF) ablation remains limited, primarily due to inter-patient variability in AF mechanisms. The ventricular residuals in ECG f-wave extraction, along with the low temporal resolution in Fourier spectral analysis, significantly impact dynamic structure analysis and may compromise the accuracy of AF recurrence prediction. To address these challenges, this work aims to improve the interpretation of recurring patterns in AF cycle length (AFCL) to aid in preoperative patient screening.</div></div><div><h3>Methods:</h3><div>The study utilized data from a dataset of 87 patients (77 with persistent AF and 10 with paroxysmal AF). The variability of AFCL was derived from the extracted f-waves of lead V1 in preprocedural 250-second recordings with EEMD-based cycle identification. Recurrence plot indices (RPIs) from recurrence quantification analysis were introduced to characterize the dynamic structure of AFCL variability. A support vector machine prediction model was subsequently applied in 10-fold cross-validation to incorporate multivariate RPIs with feature selection.</div></div><div><h3>Results:</h3><div>RPIs showed significant differences between recurrence and non-recurrence patients. In ten-fold cross-validation, the sensitivity, specificity and accuracy of the prediction model were 75%, 100%, 90% for paroxysmal AF, and 66%, 75%, 71% for persistent AF. The recurrence prediction indicated significant differences in AF-free likelihood between patients predicted to recur and those predicted not, yielding p-values of 0.004 for paroxysmal AF and 0.001 for persistent AF.</div></div><div><h3>Conclusion:</h3><div>Non-invasive AFCL dynamics analysis showed effective prediction of long-term outcomes, suggesting their potential to aid in patient selection for optimal AF ablation benefits and reveal recurrence-related AF mechanisms.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142438418","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":"Dysphagia screening with sEMG, accelerometry and speech: Multimodal machine and deep learning approaches","authors":"","doi":"10.1016/j.bspc.2024.107030","DOIUrl":"10.1016/j.bspc.2024.107030","url":null,"abstract":"<div><div>Dysphagia is a swallowing disorder that affects food, liquid, or saliva transit from the mouth to the stomach. Dysphagia leads to malnutrition, dehydration, and aspiration of the bolus into the respiratory system, which can lead to pneumonia with subsequent death. Clinically accepted dysphagia diagnosis and follow-up methods are invasive, uncomfortable, expensive, and experience-dependent. This paper explores a multimodal non-invasive approach to objectively assess dysphagia with three biosignals: surface electromyography, accelerometry-based cervical auscultation, and speech. The defined acquisition protocol was applied to patients with dysphagia and healthy control subjects. Features were extracted from the three biosignals in different domains with the aim of proposing interpretable biomarkers. Finally, the methodology was evaluated according to the accuracy and area under the receiver operating characteristic curve obtained with different classifiers. According to our results, all signals demonstrated their suitability for dysphagia screening, specially speech and multi-modal scenarios evaluated with machine learning models and also with Gated Multimodal Units. This paper contributes to reducing the knowledge gap about swallowing-related phenomena and incorporates non-invasive and multi-modal methods with high potential to be transferred and implemented in clinical practice.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142438419","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":"Study on multidimensional emotion recognition fusing dynamic brain network features in EEG signals","authors":"","doi":"10.1016/j.bspc.2024.107054","DOIUrl":"10.1016/j.bspc.2024.107054","url":null,"abstract":"<div><div>Accurate emotion recognition is crucial in scientific research and has widespread applications in medical and educational fields. Existing studies have explored to some extent the effects of space, time, and frequency domain features of EEG signals on emotion recognition but have neglected the extensive spatial features contained in the complex information interactions reflected in the synergistic work between different brain regions, as well as the effects of the interactions between the time–frequency-space features on the fusion of the dynamic features of the captured emotional state. To tackle these challenges, this paper presents a multi-dimensional emotion recognition method that incorporates dynamic brain functional network features of EEG signals. The method analyzes spatial connectivity patterns associated with emotional representations by constructing a dynamic brain functional network, aiming to capture time–space features in the EEG signals; Simultaneously, time–frequency feature extraction is achieved by using time–frequency map transformation to fine-tune the pre-trained ResNet18 model. DE and PSD of each frequency band are extracted as complementary frequency domain features through frequency band segmentation. This paper also proposes a bidirectional long and short-term memory network that incorporates an improved attention mechanism to fuse time–frequency-spatial features and consider interactions between multidimensional features. The recognition accuracies on the arousal dimension and valence dimension on the DEAP dataset reached 97.01% and 94.92%, respectively. The recognition accuracy on the SEED dataset reached 92.97%. This fully proves that the emotion recognition method described in this paper effectively extracts multidimensional features, leading to a significant improvement in emotion recognition accuracy.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":null,"pages":null},"PeriodicalIF":4.9,"publicationDate":"2024-10-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142437769","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}