The contrast of white colposcopy images is low, which is not conducive to the computer assisted identification of different degrees of diseased tissue. In order to improve the sampling accuracy under the image guidance of colposcopy, in this paper, we propose a Computer-aided cervical cancer screening method based on Multi-spectral Narrow-Band Imaging (CMNBI). We sequentially get images of cervical tissue under different illumination sources including white light, narrow-band blue light at a center wavelength of 415nm, and narrow-band green light at a center wavelength of 540nm. The multi-spectral pathology diagnosis methods consist of two stages: the first one is image preprocessing and the other is tissue classification. The image preprocessing algorithm consists of the following steps: First, we perform filtering process on three modes of images to remove noises. Secondly, the sequentially obtained images are spatially co-registered. Thirdly, the multiple narrow-band spectral images are fused. In the stage of tissue classification, a two-class K-means clustering algorithm is used, using clinics manually identified diseased region as the seed points. To eliminate strong specular reflection points of cervical tissue, we then applied improved K-means clustering algorithm combined with contour coefficient method to improve robustness of the proposed computer-aided cervical cancer screening method. To evaluate the proposed method, we apply the method to both the fused narrow-band multispectral images as well as the conventional white light images. As a result, the sensitivity, specificity and accuracy of CMNBI are all improved with the fused narrow-band multispectral images over that of the conventional white light images.
{"title":"Computer-aided Cervical Cancer Screening Method based on Multi-spectral Narrow-band Imaging","authors":"Zihan Yang, Dingrong Yi, Jiahao Shen","doi":"10.1145/3354031.3354037","DOIUrl":"https://doi.org/10.1145/3354031.3354037","url":null,"abstract":"The contrast of white colposcopy images is low, which is not conducive to the computer assisted identification of different degrees of diseased tissue. In order to improve the sampling accuracy under the image guidance of colposcopy, in this paper, we propose a Computer-aided cervical cancer screening method based on Multi-spectral Narrow-Band Imaging (CMNBI). We sequentially get images of cervical tissue under different illumination sources including white light, narrow-band blue light at a center wavelength of 415nm, and narrow-band green light at a center wavelength of 540nm. The multi-spectral pathology diagnosis methods consist of two stages: the first one is image preprocessing and the other is tissue classification. The image preprocessing algorithm consists of the following steps: First, we perform filtering process on three modes of images to remove noises. Secondly, the sequentially obtained images are spatially co-registered. Thirdly, the multiple narrow-band spectral images are fused. In the stage of tissue classification, a two-class K-means clustering algorithm is used, using clinics manually identified diseased region as the seed points. To eliminate strong specular reflection points of cervical tissue, we then applied improved K-means clustering algorithm combined with contour coefficient method to improve robustness of the proposed computer-aided cervical cancer screening method. To evaluate the proposed method, we apply the method to both the fused narrow-band multispectral images as well as the conventional white light images. As a result, the sensitivity, specificity and accuracy of CMNBI are all improved with the fused narrow-band multispectral images over that of the conventional white light images.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116968786","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Electroencephalography (EEG) can provide a wealth of valuable information to help understand the mechanism of seizures. The automatic classification of EEG signals can help clinicians make effective judgments on whether seizures occur. In this work, a method based on combined features is proposed to classify epilepsy seizures. Firstly, discrete wavelet transform is applied to the signal, and the line length features, energy distribution proportion and approximate entropy of each sub-band signal are extracted. Then the statistical features of the raw signal are extracted, including mean, standard deviation, coefficient of variation, median absolute deviation (MAD) and interquartile range (IQR). All the features are combined and the dimension of the combined feature vector is reduced by the principal component analysis (PCA). Finally, the support vector machine (SVM) is used to classify the epileptic seizure. The dataset is from the epilepsy laboratory of the University of Bonn, Germany. The accuracy of 98.40% proves the validity of this method.
{"title":"Epileptic Seizure Classification based on the Combined Features","authors":"Jie Yu, Lirong Wang, Xueqin Chen","doi":"10.1145/3354031.3354054","DOIUrl":"https://doi.org/10.1145/3354031.3354054","url":null,"abstract":"Electroencephalography (EEG) can provide a wealth of valuable information to help understand the mechanism of seizures. The automatic classification of EEG signals can help clinicians make effective judgments on whether seizures occur. In this work, a method based on combined features is proposed to classify epilepsy seizures. Firstly, discrete wavelet transform is applied to the signal, and the line length features, energy distribution proportion and approximate entropy of each sub-band signal are extracted. Then the statistical features of the raw signal are extracted, including mean, standard deviation, coefficient of variation, median absolute deviation (MAD) and interquartile range (IQR). All the features are combined and the dimension of the combined feature vector is reduced by the principal component analysis (PCA). Finally, the support vector machine (SVM) is used to classify the epileptic seizure. The dataset is from the epilepsy laboratory of the University of Bonn, Germany. The accuracy of 98.40% proves the validity of this method.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134082693","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The weighted integrated multi-scale entropy(WIMSE) is analyzed for the 135 about ten minutes RR interval series preceding the onset of ventricular tachycardia and ventricular fibrillation(called VT/VF series),and the change of WIMSE is discussed for the data samples of significant increase of heart rate (called SI_HR group) and no significant change of heart rate (called nSI_HR group) preceding the onset of VT/VF events. Results show that the WIMSE of VT/VF series has significantly reduction compared with normal sinus rhythm (scale:1--30, p<0.05),and the reduction of WIMSE is more significant for the VT/VF series of SI_HR group, the extracted complexity index (scale:1-10, p<10-6). Therefore the WIMSE may be an effective nonlinear predictive parameters for forecasting VT/VF events.
{"title":"Forecasting of Ventricular Tachyarrhythmia based on Multi-scale Entropy of Short-term Heart Rate Variability","authors":"L. Qing, D. Hong-sheng, Ma Yin-yuan","doi":"10.1145/3354031.3354048","DOIUrl":"https://doi.org/10.1145/3354031.3354048","url":null,"abstract":"The weighted integrated multi-scale entropy(WIMSE) is analyzed for the 135 about ten minutes RR interval series preceding the onset of ventricular tachycardia and ventricular fibrillation(called VT/VF series),and the change of WIMSE is discussed for the data samples of significant increase of heart rate (called SI_HR group) and no significant change of heart rate (called nSI_HR group) preceding the onset of VT/VF events. Results show that the WIMSE of VT/VF series has significantly reduction compared with normal sinus rhythm (scale:1--30, p<0.05),and the reduction of WIMSE is more significant for the VT/VF series of SI_HR group, the extracted complexity index (scale:1-10, p<10-6). Therefore the WIMSE may be an effective nonlinear predictive parameters for forecasting VT/VF events.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132959739","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In this paper we discuss the problem of automatic contour extraction of facial spot based on RGB images. Prior similar work has been frequently used for processing those hyperpigmentation skin conditions such as melasma and melanoma, where the separation between pigmented area and normal skin is easy to define. However the melanin under facial spots is normally deposited in a scatter way and distributed superficially, this makes the contrast between the area of spots and that of normal skin become small. As such it is difficult to directly extract the contour of the spots. After analyzing the individual three color channels of facial spot RGB skin image, we found that the blue channel provides the clearest edge of the spots, while the edge presents a certain amount of blur in the red channel. Therefore, this study proposed a new image processing strategy for facial spots analysis, i.e. to firstly separate the RGB channels to obtain the blue channel, then, the maximum entropy threshold segmentation and the Snake method are used to extract the contour of color spots. The experiments verified that the separated color channel and Snake-based method can help to reliably extract edge contours and preserve the color information of the spot.
{"title":"Facial Spot Contour Extraction based on Color Image Processing","authors":"Xiaojin Liu, Jiuai Sun, Xiong Wang","doi":"10.1145/3354031.3354043","DOIUrl":"https://doi.org/10.1145/3354031.3354043","url":null,"abstract":"In this paper we discuss the problem of automatic contour extraction of facial spot based on RGB images. Prior similar work has been frequently used for processing those hyperpigmentation skin conditions such as melasma and melanoma, where the separation between pigmented area and normal skin is easy to define. However the melanin under facial spots is normally deposited in a scatter way and distributed superficially, this makes the contrast between the area of spots and that of normal skin become small. As such it is difficult to directly extract the contour of the spots. After analyzing the individual three color channels of facial spot RGB skin image, we found that the blue channel provides the clearest edge of the spots, while the edge presents a certain amount of blur in the red channel. Therefore, this study proposed a new image processing strategy for facial spots analysis, i.e. to firstly separate the RGB channels to obtain the blue channel, then, the maximum entropy threshold segmentation and the Snake method are used to extract the contour of color spots. The experiments verified that the separated color channel and Snake-based method can help to reliably extract edge contours and preserve the color information of the spot.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124424554","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Enhancers are sequences in the genome that regulate gene expression and are usually located far from transcription start sites. Enhancers regulate gene expression by interacting with promoters. Therefore, the recognition of the association between enhancers and promoters is an important issue in the study of enhancer regulation. At present, computational methods to recognize the association between enhancers and promoters are mainly realized by designing machine learning methods based on the biological signals on the genome sequence. These recognition methods ignore evaluating the classification power of features, resulting in limited recognition performance. In this paper, the classification power of the feature signals near enhancers and promoters in the genome sequence was evaluated, and the features with strong classification power were picked up. This was conducive to improving the recognition accuracy. The correlation between enhancers and promoters was recognized by the random forest method. Compared with the five main recognition methods, the accuracy of the recognition method in this paper is higher.
{"title":"An Approach for Recognition of Enhancer-promoter Associations based on Random Forest","authors":"Tianjiao Zhang, Yadong Wang","doi":"10.1145/3354031.3354039","DOIUrl":"https://doi.org/10.1145/3354031.3354039","url":null,"abstract":"Enhancers are sequences in the genome that regulate gene expression and are usually located far from transcription start sites. Enhancers regulate gene expression by interacting with promoters. Therefore, the recognition of the association between enhancers and promoters is an important issue in the study of enhancer regulation. At present, computational methods to recognize the association between enhancers and promoters are mainly realized by designing machine learning methods based on the biological signals on the genome sequence. These recognition methods ignore evaluating the classification power of features, resulting in limited recognition performance. In this paper, the classification power of the feature signals near enhancers and promoters in the genome sequence was evaluated, and the features with strong classification power were picked up. This was conducive to improving the recognition accuracy. The correlation between enhancers and promoters was recognized by the random forest method. Compared with the five main recognition methods, the accuracy of the recognition method in this paper is higher.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116604976","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Wenqiang Cai, Lishen Qiu, Wanyue Li, Jie Yu, Lirong Wang
In order to improve the accuracy and efficiency of the fall detection, we proposed a fall detection algorithm based on Adaboost with single-layer decision tree under six-axis acceleration (three-axis acceleration, three-axis angular acceleration) time features. We set two thresholds for the resultant linear acceleration. When the value of the resultant linear acceleration is within these two thresholds, the algorithm of fall detection classifier is triggered. The fixed window is used to intercept the time waveform of the six-axis acceleration and extract the time features. We selected seven features with less computational complexity, and finally used these seven features to construct a fall detection model based on Adaboost with single-layer decision tree. Our algorithm can achieve 99.08% accuracy in the data set collected by ourselves, and has high specificity and sensitivity. The most critical point is that the algorithm proposed in this paper has a small computational cost and can be transplanted onto the embedded system, which is a practical and reliability fall detection method.
{"title":"Practical Fall Detection Algorithm based on Adaboost","authors":"Wenqiang Cai, Lishen Qiu, Wanyue Li, Jie Yu, Lirong Wang","doi":"10.1145/3354031.3354056","DOIUrl":"https://doi.org/10.1145/3354031.3354056","url":null,"abstract":"In order to improve the accuracy and efficiency of the fall detection, we proposed a fall detection algorithm based on Adaboost with single-layer decision tree under six-axis acceleration (three-axis acceleration, three-axis angular acceleration) time features. We set two thresholds for the resultant linear acceleration. When the value of the resultant linear acceleration is within these two thresholds, the algorithm of fall detection classifier is triggered. The fixed window is used to intercept the time waveform of the six-axis acceleration and extract the time features. We selected seven features with less computational complexity, and finally used these seven features to construct a fall detection model based on Adaboost with single-layer decision tree. Our algorithm can achieve 99.08% accuracy in the data set collected by ourselves, and has high specificity and sensitivity. The most critical point is that the algorithm proposed in this paper has a small computational cost and can be transplanted onto the embedded system, which is a practical and reliability fall detection method.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131289456","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Many machine-learning methods have been widely applied to predict Alzheimer's disease based on functional magnetic resonance imaging (fMRI) data. In our previous study, we proposed the Euler Elastica Regularized Logistic Regression (EELR) method and demonstrated its advantages over the other classifiers. In this study, we applied EELR to resting-state fMRI (RS-fMRI) data of 24 healthy aged subjects and 22 Alzheimer's disease (AD) patients for the identification of Alzheimer's disease. Moreover, in order to reveal the neural discriminative pattern, permutation test was performed to test the differences of EELR weight between AD and healthy aged subject. The results showed that EELR classifier could successfully classify AD and healthy aged subject. Moreover, EELR revealed that the amplitude of low-frequency fluctuations (ALFF) of posterior cingulate cortex, prefrontal cortex and hippocampus are the important biomarkers for distinguishing AD and healthy aged subject.
{"title":"Application of Euler Elastica Regularized Logistic Regression on Resting-state fMRI for Identification of Alzheimer's Disease","authors":"W. Guo, L. Yao, Zhi-ying Long","doi":"10.1145/3354031.3354036","DOIUrl":"https://doi.org/10.1145/3354031.3354036","url":null,"abstract":"Many machine-learning methods have been widely applied to predict Alzheimer's disease based on functional magnetic resonance imaging (fMRI) data. In our previous study, we proposed the Euler Elastica Regularized Logistic Regression (EELR) method and demonstrated its advantages over the other classifiers. In this study, we applied EELR to resting-state fMRI (RS-fMRI) data of 24 healthy aged subjects and 22 Alzheimer's disease (AD) patients for the identification of Alzheimer's disease. Moreover, in order to reveal the neural discriminative pattern, permutation test was performed to test the differences of EELR weight between AD and healthy aged subject. The results showed that EELR classifier could successfully classify AD and healthy aged subject. Moreover, EELR revealed that the amplitude of low-frequency fluctuations (ALFF) of posterior cingulate cortex, prefrontal cortex and hippocampus are the important biomarkers for distinguishing AD and healthy aged subject.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126178263","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Visual tracking is an active and challenging research topic in computer vision, as objects often undergo significant appearance variations caused by occlusion, deformation and background clutter. In recent years, many convolutional neural network based trackers have achieved impressive performance by integrating multi-layer features. However, in order to conduct multi-scale feature fusion, most of these trackers recover high-resolution presentations from low-resolution representations produced by a high-to-low resolution network, which tend to result in inaccurate feature maps or lose of details of the target object. In this paper, we propose an end-to-end region-based high-resolution fully convolutional Siamese network for tracking. In the tracker, we propose to extract the spatial information and semantic information of the target object using a high-resolution network that maintains rich high-resolution representations of the target object through the whole process. Furthermore, a set of position-sensitive score maps are obtained for all regions of the target template, and an adaptive weighting method is proposed to fuse score maps of multiple regions. Experimental results on the OTB50 and OTB100 benchmark datasets demonstrate that our tracker performs better than several state-of-the-art trackers while running in real-time.
{"title":"Region-based High-resolution Siamese Network for Robust Visual Tracking","authors":"Chunbao Li, Bo Yang","doi":"10.1145/3354031.3354051","DOIUrl":"https://doi.org/10.1145/3354031.3354051","url":null,"abstract":"Visual tracking is an active and challenging research topic in computer vision, as objects often undergo significant appearance variations caused by occlusion, deformation and background clutter. In recent years, many convolutional neural network based trackers have achieved impressive performance by integrating multi-layer features. However, in order to conduct multi-scale feature fusion, most of these trackers recover high-resolution presentations from low-resolution representations produced by a high-to-low resolution network, which tend to result in inaccurate feature maps or lose of details of the target object. In this paper, we propose an end-to-end region-based high-resolution fully convolutional Siamese network for tracking. In the tracker, we propose to extract the spatial information and semantic information of the target object using a high-resolution network that maintains rich high-resolution representations of the target object through the whole process. Furthermore, a set of position-sensitive score maps are obtained for all regions of the target template, and an adaptive weighting method is proposed to fuse score maps of multiple regions. Experimental results on the OTB50 and OTB100 benchmark datasets demonstrate that our tracker performs better than several state-of-the-art trackers while running in real-time.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123633588","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The combination of optical coherence tomography (OCT) and endoscope can take images of the body tissues for clinical diagnosis. OCT images are difficult to photograph with regular imaging devices, such as the esophagus and gastrointestinal tract. Three-dimensional reconstruction of the two-dimensional sequence images can help the doctor understand the clinical situation of the body tissue, therefore improve the accuracy of diagnosis. In this paper, Ray Casting method is used to reconstruct three-dimensional image of OCT cross-section images of guinea pig esophagus. Preprocessing including image segmentation, coordinate transformation, angle correction is used to achieve a better result in three-dimensional reconstruction. The performance of the algorithm is discussed and can achieve the same effect as what of commercial software.
{"title":"Three-dimensional Reconstruction of Optical Coherence Tomography Images of Esophagus","authors":"Sihan Nao, Miao Zhang, Lirong Wang, Yongjin Xu, Xiaohe Chen","doi":"10.1145/3354031.3354057","DOIUrl":"https://doi.org/10.1145/3354031.3354057","url":null,"abstract":"The combination of optical coherence tomography (OCT) and endoscope can take images of the body tissues for clinical diagnosis. OCT images are difficult to photograph with regular imaging devices, such as the esophagus and gastrointestinal tract. Three-dimensional reconstruction of the two-dimensional sequence images can help the doctor understand the clinical situation of the body tissue, therefore improve the accuracy of diagnosis. In this paper, Ray Casting method is used to reconstruct three-dimensional image of OCT cross-section images of guinea pig esophagus. Preprocessing including image segmentation, coordinate transformation, angle correction is used to achieve a better result in three-dimensional reconstruction. The performance of the algorithm is discussed and can achieve the same effect as what of commercial software.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130966881","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Automatic sleep staging is helpful to improve diagnosis efficiency of sleep-related diseases. This work introduces the many-to-many formulation for automatic sleep staging, which means using a many-to-many mapping to convert the contextual input to the corresponding contextual output. We use convolutional neural networks (CNNs) to perform the many-to-many mapping, and use multilayer perceptron (MLP) to merge the contextual output into the final prediction for a particular epoch. In order to avoid the influence of unobvious characteristic waves and wrong labels on the training process, this work leverages the technology of curriculum learning. By clustering algorithm based on local density, the training set is divided into several subsets according to the signal quality. We design a learning strategy by successively leveraging these subsets. To the best of our current knowledge, this is the first work using curriculum learning for automatic sleep staging. It is showed by experiments that our scheme yields an accuracy comparable to the state-of-the-art on the public dataset Sleep-EDF.
{"title":"Automatic Sleep Staging based on Curriculum Learning Approach","authors":"Xingjun Wang, Ziyao Xu","doi":"10.1145/3354031.3354033","DOIUrl":"https://doi.org/10.1145/3354031.3354033","url":null,"abstract":"Automatic sleep staging is helpful to improve diagnosis efficiency of sleep-related diseases. This work introduces the many-to-many formulation for automatic sleep staging, which means using a many-to-many mapping to convert the contextual input to the corresponding contextual output. We use convolutional neural networks (CNNs) to perform the many-to-many mapping, and use multilayer perceptron (MLP) to merge the contextual output into the final prediction for a particular epoch. In order to avoid the influence of unobvious characteristic waves and wrong labels on the training process, this work leverages the technology of curriculum learning. By clustering algorithm based on local density, the training set is divided into several subsets according to the signal quality. We design a learning strategy by successively leveraging these subsets. To the best of our current knowledge, this is the first work using curriculum learning for automatic sleep staging. It is showed by experiments that our scheme yields an accuracy comparable to the state-of-the-art on the public dataset Sleep-EDF.","PeriodicalId":286321,"journal":{"name":"Proceedings of the 4th International Conference on Biomedical Signal and Image Processing","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2019-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133324518","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}