Multi-scale features for heartbeat classification using directed acyclic graph CNN

ABSTRACT

A new architecture of deep neural networks, directed acyclic graph convolutional neural networks (DAG-CNNs), is used to classify heartbeats from electrocardiogram (ECG) signals into different subject-based classes. DAG-CNNs not only fuse the feature extraction and classification stages of the ECG classification into a single automated learning procedure, but also utilized multi-scale features and perform score-level fusion of multiple classifiers automatically. Therefore, DAG-CNN negates the necessity to extract hand-crafted features. In most of the current approaches, only the high level features which extracted by the last layer of CNN are used. Instead of performing feature level fusion manually and feeding the results into a classifier, the proposed multi-scale system can automatically learn different level of features, combine them and predict the output label. The results over the MIT-BIH arrhythmia benchmarks database demonstrate that the proposed system achieves a superior classification performance compared to most of the state-of-the-art methods.     

Optimization techniques for improving power quality data mining using wavelet packet based support vector machine

This paper aims at automatic classification of power quality events using Wavelet Packet Transform (WPT) and Support Vector Machines (SVM). The features of the disturbance signals are extracted using WPT and given to the SVM for effective classification. Recent literature dealing with power quality establishes that support vector machine methods generally outperform traditional statistical and neural methods in classification problems involving power disturbance signals. However, the two vital issues namely the determination of the most appropriate feature subset and the model selection, if suitably addressed, could pave way for further improvement of their performances in terms of classification accuracy and computation time. This paper addresses these issues through a classification system using two optimization techniques, the genetic algorithms and simulated annealing. This system detects the best discriminative features and estimates the best SVM kernel parameters in a fully automatic way. Effectiveness of the proposed detection method is shown in comparison with the conventional parameter optimization methods discussed in literature like grid search method, neural classifiers like Probabilistic Neural Network (PNN), fuzzy k-nearest neighbor classifier (FkNN) and hence proved that the proposed method is reliable as it produces consistently better results.     

Statistics over features of ECG signals

This paper presented the usage of statistics over the set of the features representing the electrocardiogram (ECG) signals. Since classification is more accurate when the pattern is simplified through representation by important features, feature extraction and selection play an important role in classifying systems such as neural networks. Multilayer perceptron neural network (MLPNN) architectures were formulated and used as basis for detection of variabilities of the ECG signals. Four types of ECG beats (normal beat, congestive heart failure beat, ventricular tachyarrhythmia beat, atrial fibrillation beat) obtained from the Physiobank database were classified. The selected Lyapunov exponents, wavelet coefficients and the power levels of power spectral density (PSD) values obtained by eigenvector methods of the ECG signals were used as inputs of the MLPNN trained with Levenberg–Marquardt algorithm. The classification results confirmed that the proposed MLPNN has potential in detecting the variabilities of the ECG signals.     

基于特征融合方法的轻微认知衰退静息态脑电数据自动检测技术研究

轻微认知衰退是阿尔茨海默病的早期阶段,而利用脑电信号进行轻微认知衰退的特征提取与分类是诊断轻微认知衰退的重要方法。在基于脑电人工智能轻微认知衰退自动检测技术中,现有研究只提取脑电波信号中的某一个特征或简单地拼接多个特征,这会导致这些方法并不能较好地考虑特征之间的相关性,并且会引发维度灾难的问题;提出了一种基于卷积神经网络的轻微认知衰退静息态脑电数据自动检测算法,通过提取脑电的功率谱及脑网络特征,并通过矩阵运算的方式对这两种特征进行融合,利用卷积神经网络对融合后的特征进行分类。该方法在上海某医院采集的数据集上获得较高的准确率;此外,通过输入特征集的不同子集,该方法找到了对轻微认知衰退最有贡献的几组特征,从而还具有一定的可解释性。在本数据集上证明了功率脑网络对于轻微认知衰退自动诊断的优势。     

基于多尺度特征提取的多导联心跳信号分类

心电图(ECG)是临床上诊断心脏疾病的重要依据,从中提取关键、有效的特征是自动诊断系统 的关键。而现今多数研究仅使用单导联或双导联数据,提取的特征不够全面,无法很好地区分不同心跳中的细 微差别。为了获得更加全面的特征和更优异的分类表现,本文提出了基于多尺度特征提取的多导联心跳信号分 类方法(MSNet)。首先,该方法接收多导联心跳信号堆叠矩阵作为输入;然后,利用 3 种不同尺度的一维卷积 分别提取特征;最后将不同尺度的特征融合并进行所属类别的分类。本文在 MIT-BIH Arrhythmia Database, MIT-BIHSupraventricular Arrhythmia Database和St Petersburg INCART 12-lead Arrhythmia Database 3个心电公开 数据集上进行了充分的实验,在五折交叉验证的策略下,对于“正常-异常”分类,该方法的准确率、召回率、 精确率、F1 值均达到了 99%以上;对于多类别分类,其平均准确率、平均召回率、平均精确率、平均 F1 值能 达到 99.5%左右。与现今优异的其他方法相比,该方法有着更好的表现。     

多模态融合的特征提取方法在SA检测中的应用

为解决睡眠呼吸暂停(Sleep Apnea, SA)检测中使用传统的机器学习方法需花大量工作在特征工程上导致效率低下,以及模型多以单通道信号进行特征提取存在识别效果不佳的问题,提出一种基于时序卷积网络(Temporal Convolutional Network, TCN)和堆叠稀疏降噪自编码器(Stacked Sparse Denoismg Auto-Encoder, SSDAEs)的多模态特征融合模型来实现特征自动提取。该模型以心电和呼吸2种信号作为输入,首先利用TCN网络提取输入信号的时序特征,然后通过SSDAEs提取信号的浅层与深层的高维特征,对于不同特征空间的心电信号特征和呼吸信号特征采用一个小型神经网络进行特征融合,将该模型与随机森林算法结合,用于解决SA片段检测问题。实验结果表明,该方法在SA片段检测的准确率、灵敏度、特异性分别是91.5%、88.9%、90.8%。通过与以往相关研究对比,验证了该模型的SA检测性能更好,效率更高。     

Wavelet transform-based feature extraction for ultrasonic flaw signal classification

In this paper, we present automatic classification models for ultrasonic flaw signals acquired from carbon-fiber-reinforced polymer specimens. Different state-of-the-art strategies based on wavelet transform are utilized for feature extraction. Furthermore, a wavelet packet transform-based local energy feature extraction method is proposed to solve the deficiencies of the existing methods. Artificial neural networks and support vector machines are trained to validate the effectiveness of different feature extraction methods for flaw signal classification. Experimental results show that the proposed method can extract reliable features to effectively classify the different ultrasonic flaw signals with high accuracy.     

Detection of mitotic cells in breast cancer histopathological images using deep versus handcrafted features

One of the most important processes in the diagnosis of breast cancer, which is the leading mortality rate in women, is the detection of the mitosis stage at the cellular level. In literature, many studies have been proposed on the computer-aided diagnosis (CAD) system for detecting mitotic cells in breast cancer histopathological images. In this study, comparative evaluation of conventional and deep learning based feature extraction methods for automatic detection of mitosis in histopathological images are focused. While various handcrafted features are extracted with textural/spatial, statistical and shape-based methods in conventional approach, the convolutional neural network structure proposed on the deep learning approach aims to create an architecture that extracts the features of small cellular structures such as mitotic cells. Mitosis detection/counting is an important process that helps us assess how aggressive or malignant the cancer’s spread is. In the proposed study, approximately 180,000 non-mitotic and 748 mitotic cells are extracted for the evaluations. It is obvious that the classification stage cannot be performed properly due to the imbalanced numbers of mitotic and non-mitotic cells extracted from histopathological images. Hence, the random under-sampling boosting (RUSBoost) method is exploited to overcome this problem. The proposed framework is tested on mitosis detection in breast cancer histopathological images dataset provided from the International Conference on Pattern Recognition (ICPR) 2014 contest. In the results obtained with the deep learning approach, 79.42% recall, 96.78% precision and 86.97% F-measure values are achieved more successfully than handcrafted methods. A client/server-based framework has also been developed as a secondary decision support system for use by pathologists in hospitals. Thus, it is aimed that pathologists will be able to detect mitotic cells in various histopathological images more easily through necessary interfaces.

     

基于对抗域适应的心电信号深度学习分类算法

心血管疾病已成为威胁人类生命健康的主要疾病之一。心电图是临床上常见的诊断心律失常的重要方法并被广泛用于监测心脏病患者的健康状况。由于现有的医疗资源匮乏,使用人工智能的方法来分析和诊断从而克服这些局限性的需求愈加迫切,在临床中使用自动检测和分类方法,可以帮助医生对疾病做出准确、快速的诊断。本文对8种常见的心律失常类型进行分类,提出一种基于对抗域自适应的心电信号深度学习分类的方法,解决并改善训练样本标注不足和个体差异导致的数据分布差异现象等问题。该方法包括3个模块：多尺度特征提取A模块、域识别B模块和多分类器C模块。A模块由2组不同的并行卷积块组成,增加了特征提取的宽度。B模块由3个卷积块和1个全连接层组成,用于充分提取浅层特征。在C模块中,将时间特征和深度学习提取特征串联在全连接层上,增强特征多样性。实验结果表明,该方法在准确率、敏感性和阳性预测值上可达到98.8%、97.9%和98.1%,所提出的模型可以帮助医生在常规心电图中准确地检测不同类别的心律失常。     

Feature saliency using signal-to-noise ratios in automated diagnostic systems developed for ECG beats

Artificial neural networks (ANNs) have been used in a great number of medical diagnostic decision support system applications and within feedforward ANNs framework there are a number of established measures such as saliency measures for identifying important input features. By identifying a set of salient features, the noise in a classification model can be reduced, resulting in more accurate classification. In this study, a signal-to-noise ratio (SNR) saliency measure was employed to determine saliency of input features of multilayer perceptron neural networks (MLPNNs) used in classification of electrocardiogram (ECG) beats (normal beat, congestive heart failure beat, ventricular tachyarrhythmia beat, atrial fibrillation beat) obtained from the Physiobank database. The SNR saliency measure determines the saliency of a feature by comparing it to that of an injected noise feature and the SNR screening method utilizes the SNR saliency measure to select a parsimonious set of salient features. ECG signals were decomposed into time–frequency representations using discrete wavelet transform. Input feature vectors were extracted using statistics over the set of the wavelet coefficients. The MLPNNs used in the ECG beats-classification were trained for the SNR screening method. The application results of the SNR screening method to the ECG signals demonstrated that classification accuracies of the MLPNNs with salient input features are higher than that of the MLPNNs with salient and non-salient input features.     

一种基于压缩域的穿戴式心电信号直接特征提取与分类方法

压缩感知是实现可穿戴式健康监测系统低能耗工作方式的一种有效途径,而现有基于压缩感知的心电信号分类方法大多需要在进行分类之前,先使用重构算法恢复出原始心电信号,这可能会导致较高的计算复杂度高,不适合于具有实时性需求的可穿戴式系统。提出一种基于压缩域的穿戴式心电信号的特征提取与自动分类方法。跳过信号重构步骤,使用改进的主成分分析法在压缩域上直接对压缩后的心电信号进行特征提取,并基于最小二乘支持向量机半监督学习方法实现心电信号的自动分类。实验结果表明,相较于在非压缩域上的分类方法,该方法在保证分类性能下降非常少的前提下,心电数据量大大地减少,有效提高了心电信号自动分类的效率。     

深度稀疏自编码器在ECG特征提取中的应用

针对心电（ECG）信号智能分析模型中,复杂波形的特征提取困难,人工设计特征造成源信号特征丢失,标签样本不足等问题,提出了一种基于深度稀疏自编码器（Deep Sparse Auto-Encoders,DSAEs）的ECG特征提取方法。该方法在DSAEs进行贪婪逐层训练时,采用适应性矩阵估计（Adaptive moment estimation,Adam）对网络权重进行寻优,以此获得最优参数组合,同时提取出高层隐含层的输出,并作为ECG高度抽象的低维特征。最后利用支持向量机（Support Vector Machines,SVM）构建分类模型,完成对ECG的特征分类。使用MIT-BIH心律失常数据库的ECG数据进行仿真实验,结果表明,提出的ECG特征提取方法能有效地分层抽取特征,提高分类识别准确率。     

基于U-NET网络的心律失常信号识别算法研究

在心电信号心律失常自动识别系统中,针对心电信号形态复杂导致特征提取困难、自动分类模型准确度低、现实应用性差的问题,设计了一种基于U-NET全卷积神经网络的心电信号语义分割的识别分类方法。该方法通过全卷积神经网络的编码运算规则,将心电信号切片数据作为输入,标签地图作为输出,可划分出信号片段中的心拍位置与类别。仿真结果表明：该方法在正常窦性搏动、左束支传导阻滞、右束支传导阻滞、房性早搏和室性早搏五分类问题中取得较高准确率,实现了对心律失常信号的有效识别。     

基于统计特征和熵特征融合的心肌梗死辅助诊断方法

针对心肌梗死临床诊断过程中临床实用性和准确率不高的问题,提出一种基于12导联心电图（ECG）的心肌梗死的辅助诊断方法。首先,对12导联ECG信号进行去噪和数据增强处理;其次,分别对各导联ECG信号提取包含标准差、峰度系数、偏度系数的统计特征,以此反映信号的形态特征;同时,提取包含香农熵、样本熵、模糊熵、近似熵和排列熵的熵特征,以此表征ECG信号时间序列的时间与频谱复杂性、新模式产生的概率、规律性和不可预测性以及检测ECG信号的微小变化;然后,融合ECG信号的统计特征和熵特征;最后,基于随机森林算法在病人内和病人间两种模式下对算法进行分析和验证,并通过交叉验证防止过拟合。实验结果表明,病人内模式下算法准确率和F1值分别为99.98%和99.99%,病人间模式下算法准确率和F1值分别为94.56%和97.05%;与基于单导联ECG的诊断方法相比,采用12导联ECG诊断心肌梗死更符合医生临床诊断逻辑。     

Signal‐to‐noise ratios for measuring saliency of features extracted by eigenvector methods from ophthalmic arterial Doppler signals

Abstract: Features are used to represent patterns with minimal loss of important information. The feature vector, which is composed of the set of all features used to describe a pattern, is a reduced‐dimensional representation of that pattern. Medical diagnostic accuracies can be improved when the pattern is simplified through representation by important features. By identifying a set of salient features, the noise in a classification model can be reduced, resulting in more accurate classification. In this study, a signal‐to‐noise ratio saliency measure was employed to determine the saliency of input features of recurrent neural networks (RNNs) used in classification of ophthalmic arterial Doppler signals. Eigenvector methods were used to extract features representing the ophthalmic arterial Doppler signals. The RNNs used in the ophthalmic arterial Doppler signal classification were trained for the signal‐to‐noise ratio screening method. The application results of the signal‐to‐noise ratio screening method to the ophthalmic arterial Doppler signals demonstrated that classification accuracies of RNNs with salient input features are higher than those of RNNs with salient and non‐salient input features.     

Semantic segmentation based on fusion of features and classifiers

This paper proposes a feed forward architecture algorithm using fusion of features and classifiers for semantic segmentation. The algorithm consists of three phases: Firstly, the features from hierarchical convolutional neural network (CNN) and the features based on region are extracted and fused on super pixel level; secondly, multiple classifiers of Softmax, XGBoost and Random Forest are ensemble to compute the per-pixel class probabilities; at last, a fully connected conditional random field is employed to enhance the final performance. The hierarchical features contain more global evidence and the region features contain more local evidence. So the fusion of these two features is expected to enhance the feature representation ability. In classification phase, integrating multiple classifiers aims to improve the generalization ability of classification algorithms. Experiments are conducted on Sift-Flow datasets by our proposed methods with competitive labeling accuracy.

     

Feature clustering and ranking for selecting stable features from high dimensional remotely sensed data

ABSTRACT

High dimensional remote sensing data sets typically contain redundancy amongst the features. Traditional approaches to feature selection are prone to instability and selection of sub-optimal features in these circumstances. They can also be computationally expensive, especially when dealing with very large remote sensing data sets. This article presents an efficient, deterministic feature ranking method that is robust to redundancy. Affinity propagation is used to group correlated features into clusters. A relevance criterion is evaluated for each feature. Clusters are then ranked based on the median of the relevance values of their constituent features. The most relevant individual features can then be selected automatically from the best clusters. Other criteria, such as computation time or measurement cost, can optionally be considered interactively when making this selection. The proposed feature selection method is compared to competing filter approach methods on a number of remote sensing data sets containing feature redundancy. Mutual information and naive Bayes relevance criteria were evaluated in conjunction with the feature selection methods. Using the proposed method it was shown that the stability of selected features improved under different data samplings, while similar or better classification accuracies were achieved compared to competing methods.     

Automatic microstructural characterization and classification using dual tree complex wavelet-based features and Bees Algorithm

During the gas tungsten arc welding of nickel-based superalloys, the secondary phases such as Laves and carbides are formed in final stage of solidification. But, other phases such as γ″ and δ phases can precipitate in the microstructure, during aging at high temperatures. However, it is possible to minimize the formation of the Nb-rich Laves phases and therefore reduce the possibility of solidification cracking by adopting the appropriate welding conditions. This paper aims at the automatic microstructurally characterizing the kinetics of phase transformations on an Nb-base alloy, thermally aged at 650 and 950 °C for 10, 100 and 200 h, through backscattered ultrasound signals at frequency of 4 MHz and dual tree complex wavelet transform (DTCWT)-based feature extraction technique. The feature set comprises of statistical attributes (such as variance, skewness and kurtosis) extracted from the complex wavelet coefficients which are obtained using the DTCWT decomposition of a backscattered ultrasound signal. Also, the performance of the proposed feature extraction technique is compared with the conventional discrete wavelet transform. Finally, these features are fed to the probabilistic neural network (PNN) and radial basis function classifiers to automatic microstructural classification. The training process of these networks depends on the selection of the smoothing parameter of the networks’ activation function in a hidden layer. In this article, we introduce the application of the Bees Algorithm to the automatic adaptation of smoothing parameters. The proposed feature extraction technique coupled with the optimized PNN yielded the highest average accuracy of 96 and 83 %, respectively, for thermal aging at 650 and 950 °C. Thus, the proposed processing system provides high reliability to be used for microstructure characterization through ultrasound signals.

     

基于半监督特征融合的监控视频场景识别研究

针对单模态特征条件下监控视频的场景识别精度与鲁棒性不高的问题,提出一种基于特征融合的半监督学习场景识别系统。系统模型首先通过卷积神经网络预训练模型分别提取视频帧与音频的场景描述特征;然后针对场景识别的特点进行视频级特征融合;接着通过深度信念网络进行无监督训练,并通过加入相对熵正则化项代价函数进行有监督调优;最后对模型分类效果进行了仿真分析。仿真结果表明,上述模型可有效提升监控场景分类精度,满足针对海量监控视频进行自动化结构化分析等公安业务需求。     

Emotion recognition based on sparse representation of phase synchronization features

Emotion recognition based on Electroencephalogram (EEG) has attracted much attention in brain-computer interaction. However, most existing methods usually focus on amplitude and spectrum of the EEG signal, leading to sub-optimal performances due to the insufficiency in modelling the complex intrinsic information of neural integration. To address this issue, this paper proposes to capitalize on the largely neglected phase synchronization (PS) between EEG channels which reflects the intrinsic rhythmic interactions between different channels in neural integration. Specifically, this paper develops a simple and novel feature extraction method which calculates the PS based sparse representation features to analyze emotion states. First, the EEG phase synchronization indexes (PSI) of all channel pairs are estimated as features to distinguish different emotions, since certain topographical maps on PSI reveal specific emotion states. Then principal component analysis is performed to eliminate redundant and noisy features in PSI. Finally, Sparse Representation based Classification (SRC) furtherly emphasize emotion-related features and restrain useless features. For the benchmark affective EEG dataset DEAP, the proposed method based on no-overlapping EEG features achieve an average accuracy of 94.5%, 87.61%, and 67.04% for the classification tasks respectively on two, three and four emotions, demonstrating the superiority over state-of-the-art emotion classification methods.