【作者】 张瑞华；

【导师】 王枞；

【作者基本信息】 北京邮电大学 ， 软件工程， 2023， 博士

【摘要】 随着医疗行业的发展和智慧医疗的兴起,基于人工智能的计算机辅助诊断相关研究也受到广泛关注。医学图像分析可以用于疾病的诊断、治疗的评测以及传染性疾病的预警等。胸部疾病分类是医学图像分析的核心研究方向之一,其旨在针对胸部的医学图像进行疾病分类。而目前应用最广泛的胸部医学图像是胸部X射线。对胸部X射线图像进行准确、快速的疾病分类以辅助医生决策,具有广泛的应用前景和重要的临床意义。利用深度神经网络技术对胸部X射线图像进行胸部疾病分类是非常有挑战的任务。首先,病灶较为细微、且与特定器官强相关,但是现有模型未利用器官相关的局部细节特征。其次,病灶较为分散、且不同疾病存在关联性,而现有网络未考虑上下文特征。最后,不同域数据分布差异大、且疾病类别分布差异大,而传统无监督域自适应方法无法有效处理类别内数据差异问题。此外,推理辅助诊断系统的实现也是科研到临床应用的重要环节。针对这些挑战和问题,本文的主要研究内容和创新点如下:1.提出了一种基于多粒度表征学习的胸部疾病分类方法,使用多分支网络和注意力模块从胸部X射线图像的重要区域学习互补的全局和局部特征。该方法使用了一种多分支网络架构,在器官掩膜的引导下学习全局和局部特征用于胸部疾病分类。此外,该方法通过一种掩膜引导注意力机制网络,进一步从精准定位的全器官和单器官重要区域独立地捕获视觉线索。在此基础上,设计了一种多任务独立学习方案,通过同时优化同一疾病标签上的多个损失以最大限度地学习全局和局部特征表达。该方法使用了一个端到端的可训练深度神经网络,在大规模数据集ChestX-ray14上进行了交叉验证实验,和多种前沿的方法进行了比较。实验结果表明,该方法胸部疾病分类性能平均AUC 达到了86.51%。2.提出了一种基于上下文表征学习的胸部疾病分类方法,结合卷积神经网络和Transformer网络结构以提升模型学习疾病上下文表征的能力。该方法包含一个卷积神经网络分支用于抽取全局特征,并设计了一个Transformer分支网络通过变换窗口的自注意力机制用于捕获更多上下文相关的视觉线索,然后通过联合损失训练进行胸部疾病分类。在ChestX-ray14数据集上的实验结果表明,该方法获得了优越的胸部疾病分类性能(AUC为86.53%),在7种胸部疾病的分类性能方面,该方法优于多粒度表征学习的胸部疾病分类方法,基于上下文表征学习的方法与基于多粒度表征学习的方法相结合,分类性能AUC可达到86.81%。两种有监督范式的学习方法互补地提高了胸部疾病分类性能。3.提出了一种基于不变性表征学习的胸部疾病分类方法,旨在从胸部X射线图像数据中学习域不变且有区分性的特征表达来进行跨域胸部疾病分类。该方法使用一个端到端可训练的深度网络模型以及一种联合学习的无监督领域自适应框架,从胸部X射线图像中学习具有鲁棒性和区分性的特征表达。在此基础上,该方法联合多个损失函数用于共同监督网络的学习,包括域不变性约束、实例不变性约束和扰动不变性约束。据调研,截止到目前,本文是第一个提出同时学习域不变且有区分性特征表达的无监督领域自适应胸部疾病分类方法。该方法在常用的大规模胸部X射线图像数据集(ChestX-ray14和SYSU数据集)上进行了验证,对比模型直接跨域应用,在SYSU→ChestX-ray14任务上,跨域平均AUC提高了3.15%,在ChestX-ray14→SYSU任务上,跨域平均AUC提高了 2.53%。4.实现了胸部疾病分类推理辅助诊断系统。针对不同的应用场景和部署需求,实现了胸部X射线图像的多维度辅助诊断。设计了优化策略,对推理系统进行复杂度优化,通过实验验证了不同优化策略对模型精度和效率的影响,推理时间在NVIDIA GPU平台上提高了23.64倍。设计了Brower/Server架构的胸部疾病分类推理系统,方便不同场景的模型部署。开发了人机交互图形化界面,并对诊断环节中的重要视觉线索进行了可视化,为疾病的辅助诊断提供更多便利性和合理依据。综上所述,本文的研究内容专注于基于深度学习的胸部疾病分类方法,所提出的方法通过验证有效提高了针对胸部X射线的胸部疾病分类的性能。更多还原

【Abstract】 With the development of the medical industry and the increasing requirement for smart medical care,CAD(computer-aided diagnosis)based on artificial intelligence has received extensive attention.Medical image analysis can assist in disease diagnosis,treatment evaluation,and early warning of epidemics.Thorax disease classification aims to identify diseases based on medical images of the chest.Currently,the most widely used chest medical image is the chest X-ray(CXR).Accurate and rapid CAD based on CXR is significant for clinical applications.Thorax disease classification using deep neural networks is a challenging task.First,lesions are subtle and strongly correlated with specific organs,but existing methods do not consider in-depth the organrelated detail visual cues.Second,the lesions are scattered and correlated with different diseases.However,the existing approaches do not fully explore the contextual features.Last,different CXR datasets often have very large data shifts as well as label shifts,which may severely hinder the scalability of supervised thorax disease classification models.On the other hand,it is not trivial to directly adopt the existing unsupervised domain adaptation(UDA)methods.in the thorax disease classification task since the label shift problem will deteriorate the distribution alignment.In addition,the thorax disease classification CAD system is also very important for clinical application.To address these challenges and problems,the main research contents and innovations of this dissertation are as follows:First,this dissertation proposes a method for thorax disease classification based on multi-granularity representation learning,which is PMGAN(Part-Aware Mask-Guided Attention Network).A multi-branch network and attention modules are designed to force the network to learn complementary global and local features explicitly in an attentive manner.The multi-branch network architecture is guided by the organ mask to learn both global and local features.The mask-guided attention mechanism further searches for informative regions and visual cues within the allorgan or single-organ regions which are precisely localized by the mask.A multi-task independent learning scheme maximizes the learning of complementary global and local feature representations by simultaneously optimizing multiple losses on the same disease label.At last,PMGAN is the end-to-end trainable deep network and according to the validation experiments on the large dataset ChestX-ray 14,compared with a variety of state-of-the-art methods,PMGAN achieves thorax disease classification performance of 86.51%average AUC。Second,this dissertation proposes a method for thorax disease classification based on contextual representation learning,which is TCCNNT(Thorax Disease Classification Net Based on Joined CNN and Transformer).The method combines the advantages of CNN and Transformer to improve the classification performance.A Transformer branch network is designed to capture more contextual visual clues by shifted windows self-attention mechanism while the CNN branch is responsible for global features extraction by convolution filtering.Through joined loss training and feature fusion,the model performs thoracic disease classification.Experimental results on ChestX-ray 14 show that TC-CNNT achieves superior classification performance(AUC 86.53%).TC-CNNT outperforms PMGAN in 7 thorax disease classifications,and the combination of TC-CNNT and PGMAN can achieve the AUC performance of 86.81%.The two supervised methods can work complementarily to improve the thorax disease classification performance.Third,this dissertation proposes a method for thorax disease classification based on invariant representation learning,which is UDATC(Unsupervised Domain Adaptation-Based Thorax Disease Classification).The method aims to learn domain-invariant and discriminative feature representations from CXR data for UDA-based cross-domain thoracic disease classification.The method includes a newdesigned end-to-end trainable deep network model to learn robust and discriminative feature representations from CXR images,a UDA framework for jointly learning invariant feature representations,and a comprehensive feature learning method,which can simultaneously regularize feature representations through domain-invariant constraints,instance-invariant constraints,and perturbation-invariant constraints.To our knowledge,this is the first domain-adaptive thoracic disease classification work that simultaneously learns invariant and discriminative features.The method is validated on commonly used large-scale CXR datasets(ChestX-ray14 and SYSU).Compared with models directly crossdomain applied,on the task of SYSU→ChestX-ray14,the method exceeds by 3.15%of average AUC,and on the task of ChestX-ray14→SYSU,the method exceeds by 2.53%of average AUC.Forth,A thorax disease classification inference CAD system has been developed.An optimization strategy is proposed to decrease the complexity of the inference model.Different optimization approaches have been verified through experiments considering both accuracy and efficiency.The inference time is accelerated by 23.64 times on the NVIDIA GPU platform.A thorax disease classification inference system based on Brower/Server architecture is designed to facilitate model deployment in different scenarios.The human-computer interaction graphical interface is developed,and the important visual clues in the inference process are visualized to provide a more convenient and reasonable basis for the disease diagnosis.In summary,the research of this dissertation focuses on the deep learning-based thorax disease classification method,and the proposed methods effectively improve the performance of CXR thorax disease classification through validation.更多还原