【作者】 张天阳；

【导师】 刘江；

【作者基本信息】 中国科学院大学(中国科学院宁波材料技术与工程研究所) ， 机械制造及其自动化， 2020， 硕士

【摘要】 近年来,医学图像的自动分析的研究已经引起了广泛的关注,吸引了来自计算机、物理、材料和数学等领域的许多海内外学者。并且现有的自动诊断算法在磁共振成像(MRI)、计算机断层扫描(CT)、超声、光学相干断层扫描(OCT)等不同医学呈像模态上也有了广泛的应用。然而伴随着人工智能领域的飞速发展、医疗图像模态的增长和医学临床标记数据量少等情况,新的挑战也不断产生。例如:医学图像良好标记的稀缺性导致可用于训练人工智能算法的数据集小,而小数据集会制约训练算法的泛化性能;来自不同设备的同一模态医学图像经常被不同数量和类型的噪声污染,而这会导致在一个设备采集的数据所训练的模型通常无法在另一个设备采集的数据集上有良好的表现;还有由于像素少而导致的边界或者细小区域分割困难。然而,深度学习同时给经典的表示学习理论注入了活力,为解决上述问题带来了动力。因此本文提出了一系列基于深度学习的表示学习模型以解决医疗图像处理的几个重要问题,并在多种模态不同器官或部位的数据上得到了充分验证。首先,本文提出了一种绘画启发的生成模型(称为SkrGAN),用于估计数据分布生成高质量的医学图像。SkrGAN通过将网络推理过程分为草图引导和彩色渲染阶段,以指获取高质量的医学图像生成。实验上,本文通过视网膜彩色眼底,胸部X射线扫描,肺部计算机断层扫描(CT)和脑在磁共振成像(MRI)这四种类型的医学图像数据集和常用的评价指标对该算法进行验证。实验结果证明了该方法在医学图像合成方面获得了先进的性能。此外,本文还将SkrGAN在视网膜血管分割中作为一种数据增强方法进行应用,进一步证明了该算法的有效性。接着,本文提出了一个噪声自适应生成对抗网络(NAGAN),旨在通过将源数据集映射到目标数据集的方式来解决不同设备采集的数据集之间的差异问题。该方法在合成图像使其具有与目标域相同的噪声分布的同时,也保留了其原本的图像语义内容,从而达到减少数据集分布的差异问题。实验表明,在光学相干断层扫描(OCT)图像和超声图像上,该方法能够很好得转化图像使数据集分割或分类的结果得以改善。最后,本文建立了一个针对细小区域分割的算法。该算法首先根据形态学原理将数据的金标准切分为不同的等级,然后利用一个多通道框架分别处理不同等级的数据输出。在优化过程中,本文通过建立一个新的损失函数来度量不同等级的偏差值。实验证明,该算法在主要的三个公开视网膜血管分割数据集上目前都具有最好的性能。更多还原

【Abstract】 Recently,medical image analysis has been discussed more and more heatedly and has been a popular interdisciplinary research domain,which draws a lot attentions from scholars of many fields,such as computer sciences,physics,material sciences and mathematics.Furthermore,the existing automatic medical image analysis algorithms have also been widely used in different medical imaging modalities such as magnetic resonance imaging(MRI),computer tomography(CT),ultrasound and optical coherence tomography(OCT).However,with the rapid development of the artificial intelligence field,the growth of medical image modalities and the small amount of medical clinical marker data,new challenges continue to arise.For example:it is often difficult to obtain sufficient high quality data in clinical situation for medical image analysis establishing,which could limits the generalization performance of the designed algorithms;medical images of the same modality from different devices are often subject to different distribution,which could make the model trained on one device often fail to perform well on that of another device;and boundary or small area segmentation is also a tough case in medical image segmentation.However,deep learning has its influence on the representation learning theory,and brings higher possibility to conquer these problems.Therefore,this thesis proposes a series of deeplearning-based representation learning algorithms to solve those problems,and the proposed algorithms have been fully verified on data of different organs in various modalities.First,a painting-inspired generative model(called SkrGAN)is porposed for estimating data distribution to generate high-quality medical images.SkrGAN divides the network inference process into sketch-guided and color rendering stages to achieve high-quality medical image generation.Experimentally,this thesis validates the algorithm with four types of medical image datasets,namely,retinal color fundus,chest X-ray scan,lung computed tomography(CT),and brain in magnetic resonance imaging(MRI)with classic metrics.The experimental results prove that the method has achieved the state-of-the-art performance in medical image synthesis.In addition,this thesis also applies SkrGAN as a data augmentation method in retinal blood vessel segmentation,which further proves the effectiveness of the proposed algorithm.Then,this thesis proposes a Noise Adaptive Generation Adversarial Network(NAGAN),which aims to solve the problem of differences between data sets collected from different devices by mapping the source dataset to the target data set.The proposed method could not only make the synthesized image share the same noise distribution with that of the target domain but also retain the original image semantic content of the inputs.Therefore,NAGAN could reduce the problem caused by the domain gaps between the source domain and the target domain.Experiments show that the method can transform the image well on the optical coherence tomography(OCT)image and ultrasound image to improve the results of cross-domain segmentation or classification.Finally,this thesis establishes an algorithm for the small region segmentation.It firstly divides the ground truths of inputs into different thickness levels through morphological principles.Then it uses a multi-channel framework to obtain different-level outputs.Furthermore,a new loss function is established to measure the distance between the outputs and ground-truth.Experiments show that the algorithm reaches the state-of-the-art performance on three public retinal vascular segmentation datasets.更多还原