当前位置: 首页>博士论文>资源详情
基于卷积神经网络的生物医学信号分类与重构
中文摘要

生物医学信号是通过仪器设备可以采集到的、对于生命体状态最直接的描述信息。其种类众多且来源广泛,是生物学、信息学、医学等学科的重要研究对象。随着生物医学相关学科的迅速发展,科研人员和医务工作者对于此类信号处理、分析技术的需求愈发强烈。高效的生物信号处理方法可以有效提升科研人员探索生命机制工作的效率,从而更好地揭示生理结构与功能的关系,进而推动生物学重大发现的产生。高精度的医学信号分析策略可以在一定程度上分担医生的临床诊断工作压力,协助其制定更有利于疾病预防和治疗的方案,进而减轻患者的生理和精神痛苦,提升社会的整体健康水平。传统的信号分析方法已经很难满足日益增长的生物医学信号处理需求,利用先进的机器学习技术对所涉及问题进行有效建模和分析成为该领域的研究热点。 深度学习是指利用多层神经网络获得数据的特征表示,进而利用该特征表示进行数据分析的机器学习方法。作为当今机器学习领域最受关注的研究方向,深度学习正引领着新一轮的人工智能研究浪潮。凭借其强大的非线性特征表示能力,近年来,卷积神经网络成为众多深度学习模型中应用最广、综合效果最好的深度神经网络结构,在以模式识别为代表的诸多任务上取得了一系列前所未有的重大突破。受到卷积神经网络在其他领域成功的启发,结合具体生物医学信号的特点,本文以乳腺钼靶X光图像和脑电信号这两类典型生物医学信号为研究对象,提出了若干基于卷积神经网络的生物医学信号分类与重构方法。本文的主要创新点如下: 1.在基于钼靶X光片的乳腺肿块分类任务中,设计可以有效描述乳腺肿块视觉特性,并易于在特征空间区分良恶性肿块的视觉特征是完成高精度辅助诊断的关键。现有的手工视觉特征往往只能从单一视觉层次描述乳腺肿块的视觉特性,忽略不同层次视觉特征表示在乳腺癌诊断过程中的协同作用。单一视觉特征的表示能力有限,而传统的多特征融合方法需要经过复杂的特征选择过程。针对上述问题,本文提出了一种基于卷积神经网络特征表示的乳腺肿块良恶性分类方法。具体而言,该方法首先构造肿块图像视觉特征表示网络,并基于自然图像和乳腺图像对该网络进行训练。之后,借鉴医生的实际诊断经验,利用该网络获得适用于乳腺肿块分类的不同层次视觉特征描述。最后提出了一种乳腺肿块特征决策机制,完成对乳腺肿块良恶性的判断。实验结果表明,该算法可以有效利用卷积神经网络特征描述,实现较高精度的乳腺肿块良恶性分类。 2.在乳腺疾病的计算机辅助诊断中,经常遇到视觉特征不明显的非典型样本。在传统的手工视觉特征空间和卷积神经网络特征空间,都无法对该类型样本进行有效描述。通过特征空间转换操作,将在原始特征空间中无法被有效描述的样本映射到易于区分的特征空间是解决此类问题的有效途径。为此,针对乳腺肿块分类问题,本文提出了一种基于大间隔度量学习的改进卷积神经网络模型。具体而言,首先,通过引入大间隔度量学习损失函数,学习从原始卷积神经网络特征空间到新特征空间的映射关系。进而由该映射关系,获得类内分布更紧致、类间分布更离散的乳腺肿块特征表示。此外,通过不断向网络提供新的错误样本,提出侧重疑难病例的网络训练改进策略。实验结果表明,该算法的大间隔度量学习层部分可以提升特征的区分度和网络分类准确率,改进的网络训练策略可以进一步提升网络在良恶性乳腺肿块鉴别任务上的表现。 3.在基于脑电信号的神经解码任务中,对不同类型刺激引发的脑电信号进行分类是该领域研究的基础性工作之一。脑电信号包含丰富的时空特性,传统的脑电信号特征描述方法通常只能表示其中的一种特性,无法为后续的分类任务提供更有效的特征表示基础。针对上述问题,本文提出了一种基于时空融合卷积神经网络的脑电信号分类方法。具体而言,首先,分别通过不同的脑电激活图产生方法,生成侧重空间信息和侧重时间信息的脑电激活图。然后,分别设计并训练可以对两种不同激活图进行特征表示的卷积神经网络,用于获得脑电信号的卷积神经网络特征表示。最后,采用特征拼接和特征选择两种融合方式,实现脑电信号的分类。实验结果表明,该算法的特征表示部分可以获得具有较强区分性的脑电特征表示,后续的时空融合策略可以进一步提升脑电信号分类精度。 4.基于脑电信号的视觉刺激重构是建立在高精度脑电信号分类基础上的又一典型神经解码任务。该任务通常由高精度脑电信号特征表示和分类,以及视觉刺激生成两个阶段构成。在传统基于认知空间脑电特征的视觉刺激重构方法中,受限于人类认知水平的极限和神经信号采集过程的误差,很难在分类精度和表示效率上获得较大提升。受当前卷积神经网络在视觉任务上取得超越人类表现结果的启发,本文提出了视觉特征指导的脑电信号分类方法。该方法通过将脑电信号特征表示映射到视觉空间,实现更高精度的脑电信号分类。然后,基于视觉特征指导的脑电信号表示,本文提出了一种用于生成视觉刺激的改进生成对抗网络模型。脑电信号分类和视觉刺激生成两个阶段对应的实验结果表明,该算法可以有效提升脑电信号分类的精度和视觉刺激重构结果的质量。 关键词:生物医学信息学,深度学习,卷积神经网络,生成对抗网络,乳腺癌,神经解码,计算机辅助诊断

英文摘要

Biomedical signals are the most direct descriptions of living organisms that can be collected through instruments and devices. They usually come from a variety of sources and are widely used, and they are key research objects in biology, informatics, medicine, and other disciplines. With the rapid development of biomedical related disciplines, scientific researchers and medical workers are increasingly demanding such signal-related processing and analysis technologies. Efficient biological signal processing methods can effectively enhance researchers' exploration of life mechanisms, so as to better reveal the relationship between physiological structure and function, and then to promote the emergence of major discoveries. A high-precision medical signal analysis strategy can share the pressure of the doctors' diagnosis to a certain extent and assist them in formulating a plan that is more conducive to disease prevention and treatment, thereby reducing the patient's physical and mental pain and improving the health of the community. Traditional signal analysis methods have been difficult to meet the growing demand for biomedical signal processing. Using advanced machine learning techniques to effectively model and analyze the issues involved has become a research hotspot in this field. Deep learning refers to machine learning method using multi-layer neural networks to obtain the characteristic representation of data for further analyzing. As the focus of research in the field of machine learning today, deep learning is leading a new round of artificial intelligence trends. With its powerful nonlinear feature representation capabilities, convolutional neural networks are the most widely used and most comprehensive network structures in many deep learning models, achieving a series of unprecedented breakthroughs in many tasks represented by pattern recognition. Inspired by the success of convolutional neural networks in other fields, considering characteristics of specific biomedical signals(mammography and electroencephalogram), this paper presents a number of biomedical signal classification and reconstruction methods based on convolutional neural networks. The main innovations of this article are as follows: 1.In the classification task of breast masses based on mammography X-ray films, designing visual features that can be effectively described and easily differentiated between benign and malignant masses in the feature space is the key to complete high-precision computer-aided diagnosis. The existing manual visual features can only describe the visual characteristics of breast masses from a single visual level, and neglect the synergistic effects of different levels of visual features in the diagnosis of breast cancer. And the ability to express a single visual feature is limited. Traditional multi-feature fusion methods require complex feature selection processes. In view of the above problems, this paper proposes a breast mass classification method based on convolutional neural network feature representation. Specifically, the method first constructs and trains feature representation networks based on natural images and mammographic images. Then, referring to the doctor's actual diagnostic experience, this network was used to obtain different levels of feature descriptions suitable for the classification of breast masses. Finally, a decision mechanism for the characteristics of breast masses is proposed to complete the judgment of benign and malignant breast masses. The experimental results show that the algorithm can achieve high precision classification of benign and malignant breast masses. 2.In the computer-aided diagnosis of breast diseases, atypical samples with indistinct visual features are often difficult to be dealt with. In the traditional handcrafted visual feature space and convolutional neural network feature space, neither of these types of samples can be effectively described. Through the feature space transformation operation, mapping the samples that cannot be effectively described in the original feature space to an easily distinguishable feature space is an effective way to solve such problems. For this reason, aiming at the classification of breast masses, this paper proposes an improved convolutional neural network model based on large margin metric learning. Specifically, first, by introducing a large margin metric learning loss function, the mapping relationship from the original convolutional neural network feature space to the new feature space is learned. Furthermore, from this mapping relationship, the characteristics of masses with more compact intraclass distributions and more discrete distributions between classes are obtained. In addition, through the continuous provision of new error samples to the network, a network training improvement strategy focusing on difficult cases was proposed. Experimental results show that the large interval metric learning layer of the algorithm can improve the feature distinguishing degree and the network classification accuracy. The improved network training strategy can further improve the performance of the network in benign and malignant lump discrimination tasks. 3.In the task of neural decoding based on EEG signals, classification of EEG signals evoked by different types of stimuli is one of the basic goals in the field of research. EEG signals contain abundant spatio-temporal information. Traditional EEG signal feature description methods can only represent one of these features and cannot provide more effective feature representation for subsequent classification operations. In view of the above problems, this paper proposes an EEG signal classification method based on spatio-temporal fusion convolutional neural network. Specifically, first of all, through different generation methods of EEG activation maps, EEG activation maps focusing on spatial information and those focusing on temporal information are generated. Then, two convolutional neural networks for different kinds of activation maps is designed and trained separately to obtain a convolutional neural network feature representation of the brain electrical signals. Finally, the two methods of feature stitching and feature selection are used to achieve the classification of EEG signals. Experimental results show that the feature representation part of the algorithm can obtain differentiated EEG feature representations, and subsequent feature fusion strategies can effectively improve the classification accuracy of EEG signals. 4.The reconstruction of visual stimuli based on EEG signals is a neural decoding task based on the classification of high-precision EEG signals. This task is usually composed of highprecision EEG signal classification and feature representation, as well as visual stimulus generation in two phases. The traditional method based on EEG features in cognitive space is limited by the limit of human cognitive level and the error of neural signal acquisition, and it is difficult to obtain great improvement in classification accuracy and efficiency. Inspired by the fact that current convolutional neural networks achieve visual results beyond human performance, this paper proposes a visual feature-guided classification method of EEG signals. This method achieves higher-precision classification of EEG signals by mapping the feature representation of EEG signals into visual space. Then, based on the representation of EEG signals guided by visual features, this paper proposes an improved generational confrontation network model for generating visual stimuli. The experimental results of these two phases of EEG signal classification and visual stimuli generation show that the algorithm can effectively improve the classification accuracy of EEG signals and the quality of visual stimuli reconstruction results. Keywords: Biomedical informatics, Deep learning, Convolutional neural networks, Generative adversarial networks, Breast cancer, Neural decoding, Computer-aided diagnosis

作者相关
主题相关
看过该书的人还在看哪些书