【作者】 王炎；

【导师】 张蕾；

【作者基本信息】 四川大学 ， 计算机科学与技术， 2021， 博士

【摘要】 乳腺癌是多数国家女性人群中发病率最高的癌症,研究表明,乳腺癌的早期准确诊断并及时治疗能够大大降低其致死率。乳腺X光检查是全世界范围内使用最广泛的早期乳腺癌筛查手段之一,也是唯一被证明可以显著降低乳腺癌带致死率的医学影像学方法。使用乳腺X光进行乳腺癌筛查时,会产生一系列乳腺X光图像数据,放射医师根据这些图像数据进行良恶性诊断,而诊断结果与医生个体经验水平有着直接联系。随着计算机技术的发展,计算机辅助诊断系统开始被用于辅助医生诊断疾病。基于乳腺X光图像的计算机辅助诊断系统能够提升乳腺癌的筛查效率和准确率,极大的减轻医生工作负荷,并降低因个体因素带来的结果偏差。在基于乳腺X光图像的乳腺癌筛查辅助诊断系统中,核心部分是乳腺X光图像分类模型。乳腺X光图像分类算法也是当今医学图像分析中的研究热点和难点之一。神经网络方法是一类模拟生物神经网络信息处理过程的计算模型,具有强大的特征提取能力。研究乳腺X光图像分类的神经网络方法具有重要的理论价值和广阔的应用前景。然而现有神经网络方法在乳腺X光图像分析中存在若干不足,如现有神经网络方法对数据依赖性强,而乳腺X光图像数据的获取和标注都存在困难;及乳腺X光图像分类方法的针对性研究还有待提升。本文针对乳腺X光图像数据的特点,着重从减弱数据依赖和获取有效的特征表达两个方面研究面向乳腺X光图像分类的神经网络方法。在减弱神经网络方法的数据依赖性方面:研究多实例数据分类,减轻分类方法对每张图像标签的依赖,进而提出基于特征敏感性的神经网络模型;针对数据集规模大小对神经网络方法产生的性能影响问题,研究乳腺X光图像分类的迁移学习方法。在获取乳腺X光图像有效分类特征表达方面:研究通过优化分类任务中不同类别的类内距离和类间距离来获取可分性更高的乳腺X光图像特征,从而提升分类准确率;针对肿块分类,研究基于先定位后分类的策略,通过定位方法去除无关区域的特征,来提升乳腺肿块分类正确率。本文的创新点和主要贡献包括以下几个方面:·提出基于特征敏感度的特征融合方法,研究多实例乳腺X光图像分类方法。在基于乳腺X光的乳腺癌筛查中,病人一次检查会从多个角度拍摄多张不同视角的图像,然后由放射科医师进行诊断。在一个病人拍摄的多张图像中,若其中一张图像为恶性,则该例病人为恶性样本。对于乳腺X光图像数据,现有方法多是基于单张图像进行判断。该类方法一方面需要对每张图片都进行标注,大大增加了手工标注数据的工作量;另一方面,在智能诊断的过程中,缺少对乳腺X光图像多角度信息的使用。为了克服多实例分类的困难,本文首先收集了一个临床乳腺X光图像数据集。进一步,为了解决多实例分类问题,提出基于图像特征敏感度的特征融合方法。该方法首先对每一例病人的每张实例图像进行特征提取,然后设计一个可学习的滤波器,使用这个滤波器对一例病人中的每个实例特征进行评估。若存在恶性特征则赋予该实例数据一个较大权重,反之赋予一个较小权重。最后根据这些权重对每例病人的多个实例特征进行融合,来得到更具可分性的病例特征。在我们收集的临床数据集和一些公开数据集上进行的实验表明,本文所提出的方法优于其它类似的乳腺癌筛查方法。并且,基于临床数据的研究更具有实际临床应用价值。·提出基于对抗域适应的迁移学习方法,研究基于迁移学习的乳腺X光图像分类方法。神经网络的训练离不开大数据的支撑,而医学数据的一大难点是数据难以获取。在数据量较小时,当前常用方法的效果可能大幅度下降。而迁移学习能迁移其它数据集上学习到的知识到目标数据集上,能一定程度上缓解由于数据集小所造成的性能下降。目前在医学数据分析中,常用的迁移学习方法是直接迁移自然图像中预训练的知识,没有考虑自然图像数据和医学图像数据之间存在的特征差异,如色彩空间、物体形状等。由此,本文提出了基于域适应和迁移学习的乳腺X光图像分类方法。该方法分为两个阶段:1)使用对抗域适应网络,在公开数据集和目标数据集上进行无监督域适应学习;2)将第一阶段学习到的知识迁移到目标分类数据集上,再基于目标分类数据进行微调。第一阶段的无监督域适应方法能够从源域数据和目标域数据中学习到更接近目标数据集的特征知识,降低迁移知识在源数据和目标数据之间的差异。大量实验结果表明本文所提出的方法能够取得更好的分类性能。·提出基于多分类器约束的方法,研究乳腺X光分类问题中的类间方差和类内方差的优化方法。基于乳腺X光图像的乳腺癌筛查本质上是对乳腺X光图像进行良恶性分类。要提高分类结果就要求特征提取网络能够得到判别性较强的特征表达。本文提出一种多分类器约束方法,通过对两个分类器的决策边界进行约束,使得该方法的特征提取网络能够获得类间距离更大、类内距离更小的特征表达。同时,两个分类器能够检测分类结果不一致的样本。这些样本的特征表达是实际分布在决策边界附近的样本,分类器对这些样本的判断是困难的。通过在训练过程中对该类样本的损失函数进行加权,使得特征提取网络更加关注该类样本的特征。在乳腺X光图像公开数据集上进行实验,验证了本文方法的有效性。·提出增强学习和多任务学习的定位方法,研究对乳腺肿块先定位再分类的方法,提升肿块良恶性分类结果。针对乳腺X光图像中的肿块分类,一种常用的手段是通过定位或者分割手段,将肿块区域单独提取出来再对其进行分类。其中的目标定位方法可以分为两类,自顶向下的方法和自底向上的方法。自顶向下的方法需要从图片中逐步搜索确定出目标所在位置,而增强学习算法能够很好的实现搜索并决策这个过程。现有的基于增强学习的自顶向下方法在神经网络训练过程中,需要存储历史状态并用于训练,而存储的状态本身是一个非平衡的采样,从而导致决策过程难以学习。为了更好的学习目标定位决策过程,我们提出使用多任务学习的方法,使得网络学习过程中受到多任务学习带来的增益。同时在划分多任务时,将原有的不平衡采样打破,使少数采样类独立成一个任务,缓解了不平衡采样带来的性能瓶颈。在自然图像和乳腺X光图像数据集上实验结果表明,本文的方法更能准确的定位到目标所在位置,进而提升了分类准确率。更多还原

【Abstract】 Breast cancer is one of the most common cancers among women in most countries.Studies have shown that early precision diagnosis and timely treatment of breast cancer can save lives.Mammography is one of the most widely used methods for early breast cancer screening in the world,and also the only medical imaging method proved to significantly reduce the mortality rate of breast cancer.Mammography for breast cancer screening pro-duces several mammograms.Based on them,radiologists diagnose tumors to be benign or malignant,and the results are often highly related to the radiologists’ experience.With the development of computer technology,computer-aided diagnosis systems have been used to assist doctors in diagnosing diseases.These systems,based on mammographic image,can improve efficiency and accuracy of breast cancer screening,greatly reduce the workload of doctors,and reduce the deviation of results caused by individual factors.The key part of computer-aided diagnosis system for breast cancer screening is the classification model.Mammographic image classification is also one of the research hotspots and difficulties in medical image analysis.As a kind of computational model that simulates the information processing process of biological neural networks,the neural network method has a strong feature extraction ability.The neural network method for the classification of mammographic images has importan-t theoretical value and broad application prospects.However,there are some deficiencies in existing neural network methods in mammographic image analysis.For example,the existing neural network methods are highly data-dependent but obtaining and labeling of mammographic data is difficult.Besides,the specific research of mammographic image clas-sification methods still needs improvements.In view of the characteristics of mammographic data,this thesis focuses on the research of neural network methods for mammographic image classification from two aspects:reducing data dependence and obtaining effective feature expression.In terms of reducing the data dependence of the neural network method,we s-tudy multi-instance data classification to reduce the dependence of the classification method on the label of each image and then propose a feature sensitive neural network model.Aim-ing at reducing the effect of dataset size on the performance of neural network method,we study a transfer learning method for mammographic image classification.In terms of obtaining the effective representation for classification mammographic images,we study a method to optimize the inter-class distance and inter-class distance of different categories in the classification task to obtain higher separability representations,so as to improve the classification accuracy.For the classification of breast masses,we adopted the strategy of localization first and classification after,that is,to remove the features of irrelevant regions by the localization method to improve the classification accuracy of breast masses.The main content and contributions of this dissertation are listed as follows.·A feature sensitive feature fusion method is proposed to study the classification of multi-instance mammograms.During mammography examination,several images from multiple perspectives are tak-en for each patient at one time,which are then referred by a radiologist for diagnoses.Among these pictures,if one is malignant,then the sample of this patient is malignant.For mammographic image dataset,most of the existing methods are based on a single image.This kind of method needs each image to have a precision label,which greatly increases the workload of labeling data.Moreover,the information of multi-view images is not used.To overcome the difficulty of multi-instance classification,a dataset of clinical mammographic images was first collected.Furthermore,to solve the problem of multi-instance classifica-tion,a feature sensitive based feature fusion method is proposed.The method first extracts the features of each instance image of one patient,then designs a trainable filter,which is used to evaluate the features of each instance image.If there is a malignant feature,this instance will have a large weight,otherwise this instance will have a small weight.Finally,according to these weights,multiple instance features of each patient are fused to obtain the final features.Experiments on our collected clinical dataset and some public mammography datasets demonstrate that the proposed method is superior to other similar breast cancer screening methods.Moreover,the research based on clinical data is more significant.·An adversarial domain adaptation based transfer learning method is proposed to study the transfer learning in mammography classification.The training of the neural network method cannot be achieved without the support of big data.One big difficulty is concerned with obtaining medical data.When the scale of the dataset is small,the effect of current methods may decrease dramatically.The transfer learning can transfer the knowledge learned from other datasets to the target dataset,which can relieve the performance impact caused by small dataset to some extent.In the recent literature,the commonly used transfer learning method in mammographic image analysis is to directly transfer the knowledge of pre-training in natural images without considering the differences between natural image features and medical image features,such as color space,object shape,etc.Therefore,this thesis proposes a breast cancer screening method based on domain adaptation and transfer learning.The proposed method is divided into two stages.In the first stage,unsupervised domain adaptive learning is carried out on both the public dataset and target dataset using an adversarial domain adaptation network.In the second stage,the knowledge from the unsupervised domain adaptation stage is transferred to the target dataset,and optimization is performed on the target dataset.In the first stage,the proposed method uses unsupervised domain adaptation,which aims at learning proper features so that transferring it to the target dataset and reducing the difference between the source and target dataset.And many experiments that compare the proposed method with the current state-of-the-art methods is carried out,the experimental results show that our method has achieved better or equivalent performance.·A multiple classifiers constrain based method is propoed to study the optimizing of inter-class distance and intra-class distance in the classification problem.Breast cancer screening based on mammograms is essentially the classification of benign and malignant mammographic images.To improve the classification results,feature extrac-tion models need to extract the strong discriminative features.In this thesis,a method to optimize the decision boundary of two classifiers is proposed.By constrains the decision boundary of two classifiers,the feature extraction network part can extract the feature with a larger intra-class distance and smaller inner-class distance.Using the two classifiers,at the same time,they can detect the samples with inconsistent classification results.This kind of sample is distributed near the decision boundary,and the classifier is difficult to learn to separate.By weighting the loss function of this kind of sample during the training process,the feature extraction network pays more attention to the features of this kind of samples.Experiments on multiple public datasets demonstrate the effectiveness of the proposed method.·A reinforcement learning and multi-task learning object localization method is pro-posed.Based on localization and then classification strategy,by using the localization results to classify benign and malignant masses,which improves the classification results of benign and malignant masses.For the classification of masses in mammograms,a commonly used way is to separate out the masses by using a localization or segmentation method and then classify these masses.The object localization methods can be divided into two categories,the bottom-up and the top-down strategy.In the top-down strategy,the localization can be regarded as a decision make process.The reinforcement learning method is good at learning the process of object search and decision-making.The existing top-down strategy method based on reinforcement learning needs to store the historical states and use it for training the neural network,and the stored states itself are an imbalanced sampling,which makes the decision-making process difficult to learn.To better learn the decision-making process of localization objects,we propose a method that using a multi-task learning method,so that the neural network can make use of the advantages of multi-task learning in structure.At the same time,when dividing into multi-tasks,the original imbalanced sampling problem has been broken,and the minor sampling class is separated into one independent task,which alleviates the performance bottleneck brought by imbalanced sampling.Finally,the method in this paper was verified on natural image dataset and mammographic image dataset.The experimental results showed that the proposed method in the thesis could localize the objects more accurately.The results of benign and malignant tumor classification were improved by further apply the proposed method to the classification of mammograms.更多还原