【作者】 董阳；

【导师】 潘海为；

【作者基本信息】 哈尔滨工程大学 ， 计算机科学与技术， 2021， 硕士

【摘要】 脑肿瘤是一种在大脑内生长的异常细胞群,主要分为两种,一种是原发性脑肿瘤,另一种是转移脑肿瘤。其中胶质瘤是最为普遍的原发性脑肿瘤,具有几种不同的侵袭性和不同的组织学子区域。据最新的全国癌症报告统计,我国脑肿瘤发病率大概为十万分之七到八,年增长率虽然只有百分之一左右,但也保持着缓慢增长的趋势。脑肿瘤对人类的健康有着很大威胁,它会破坏和压迫正常的脑组织,引起脑水肿,导致颅内压增大,以致影响呼吸中枢神经等,最终导致死亡,因此脑肿瘤仍是医学界中不可忽视的一个课题。脑肿瘤的诊断过程,常利用MRI来显示脑内结构,因为可以通过增强剂等强化恶性肿瘤的亮度,能够更早更准确地确诊脑肿瘤。然而想要将脑肿瘤各个组织学子区域完整的分割出来,就需要结合MR图像多个模态的特征。脑肿瘤图像分割是脑肿瘤诊断治疗过程中的一项重要内容,当前主要依靠医生手动分割,大量的脑肿瘤图像数据给医生带来了巨大的压力,手动分割耗时耗力,并且其效果取决于医生的经验,所以研究自动分割方法对于辅助医生的诊断意义重大。目前深度卷积神经网络在医学图像语义分割方面已取得了很大的进步,但一般需要大量密集的注释图像进行训练才能取得好的效果,并且对于未知类别的泛化能力不强。然而,受仪器差异等各种因素的影响,脑肿瘤图像的私有数据集之间存在很大差异,获得大量带有监督信息的脑肿瘤图像数据非常困难。与此同时,脑肿瘤图像中脑肿瘤本身所占面积非常小,脑肿瘤像素数远小于背景类像素数,在训练的过程当中,可能会产生类别不平衡问题。针对脑肿瘤图像的这些特点,以下为本文的研究内容:(1)针对脑肿瘤MR图像的多模态特点,本文采用2D切片融合的方法,对脑肿瘤MR图像的T1,T1ce,T2,FLAIR四个模态的特征进行融合,获得包含四个模态特征的图像,为后续的分割任务做初步准备。(2)对于经过预处理的脑肿瘤MR图像,本文结合小样本学习和深度学习理论,提出一种新型网络结构PU-net(Prototype Network based on U-net),对脑肿瘤多模态MR图像进行分割。引入小样本学习的思想,能够利用有限的监督信息对模型进行泛化,以度量学习方法中的原型网络为基础进行改进,将其应用于分割任务,可以不引入额外的训练参数。(3)特征提取器在U-net的基础上进行优化,可以融合浅层特征和深层特征,能够快速学习脑肿瘤的分布规律和脑肿瘤的精确细节;对下采样方法和卷积方式进行优化,减小特征提取过程中的信息损失;利用早期融合结构提取ROI,从而避免类别不平衡问题,同时用于后续的特征提取计算。(4)在所提出的PU-net的基础上,加入注意力机制,即在特征提取器中加入注意力门,进一步提出PAU-net(Prototype Network based on U-net with Attention gate)模型对脑肿瘤进行分割,通过注意力门来提取ROI,结合后期融合替代早期融合在模型中的作用,从而进一步提升模型性能。经过试验得出,本文提出的PAU-net方法在Bra TS18数据集上的分割的各项评价指标均超越了最新的小样本分割算法,对小样本脑肿瘤多模态MR图像能实现很好的分割效果。更多还原

【Abstract】 Brain tumors are a group of abnormal cells that grow in the brain.There are two main types: primary brain tumors and metastatic brain tumors.Glioma is one of the most common primary brain tumor with several different aggressiveness and different characteristics.Histology sub-area.According to the latest statistics from the National Cancer Report,the incidence of brain tumors in China is about 7 to 8 per 100,000.Although the annual growth rate is only about 1%,it also maintains a slow growth trend.Brain tumors are a great threat to human health.They can damage and compress normal brain tissues,cause brain edema,increase intracranial pressure,and affect the respiratory central nervous system,and ultimately lead to death.Therefore,brain tumors are still a topic that cannot be ignored in the medical field.In the diagnosis process of brain tumors,MRI is often used to show the structure of the brain,because the brightness of malignant tumors can be enhanced by enhancers,etc.,which can diagnose brain tumors earlier and more accurately.However,in order to completely segment the various histological sub-regions of brain tumors,it is necessary to combine the characteristics of multiple modalities in MR images.Brain tumor image segmentation is a major content in the process of brain tumor treatment.Currently,doctors rely on manual segmentation.A large amount of brain tumor image data brings great pressure to doctors.Manual segmentation is time-consuming and labor-intensive,and its effect depends on Based on the doctor’s experience,the study of automatic segmentation methods is of great significance to assist the doctor’s diagnosis.At present,deep convolutional neural networks have made great progress in the semantic segmentation of medical images,but generally require a large number of densely annotated images for training to achieve good results,and the generalization ability for unknown categories is not strong.However,due to various factors such as instrument differences,there are big differences between private data sets of brain tumor images,and it is very difficult to obtain a large number of brain tumor image data with supervised information.At the same time,the brain tumor itself occupies a very small area in the brain tumor image,and the number of background pixels is much more than brain tumor pixels,and there is a problem of category imbalance during training.In view of these characteristics of brain tumor images,the following is the research content of this article:(1)Aiming at the multi-modal characteristics of brain tumor MR images,this paper adopts 2D slice fusion method to fuse the features of the four modalities including T1,T1 ce,T2,and FLAIR in brain tumor MR images to obtain an image containing four modal features,Make preliminary preparations for the subsequent segmentation task.(2)For pre-processed brain tumor MR images,this paper combines few-shot learning and deep learning theory,and proposes a new network structure PU-net(Prototype Network based on U-net),which is used for multi-modal MR images of brain tumors segmentation.Introducing the idea of few-shot learning,it can use limited supervision information to generalize the model,improve based on the prototype network in the metric learning method,and apply it to the segmentation task without introducing additional training parameters.(3)The feature extractor is optimized on the basis of U-net,which can fuse shallow features and deep features,and learn the distribution law of brain tumors and the precise details of brain tumors quickly;optimize the downsampling method and convolution method to reduce the information loss in the feature extraction process;using early fusion structure to extract ROI,thereby can avoid category imbalance problems,and can be used for subsequent feature extraction calculations.(4)On the basis of the proposed PU-net,an attention mechanism is added,that is,an attention gate is added to the feature extractor,and the PAU-net model is further proposed for brain tumors segmentation,and ROI is extracted through the attention gate.Combining it and late fusion instead of early fusion in the model,thereby further improving the model performance.Through experiments,it is concluded that the evaluation indexes of the segmentation of the PAU-net(Prototype Network based on U-net with Attention gate)method proposed in this paper on the Bra TS18 dataset surpass the latest few-shot segmentation algorithm,and can achieve good segmentation for few-shot brain tumor multimodal MR images effect.更多还原