毕业生知识库

图像分割是医学图像处理最常用和最重要的技术手段之一，通常作为疾病分析诊断的首要步骤。大脑作为人体最重要的器官，其相关研究对于更好理解人类思维、研究病变发展以及引导大脑发育有重要意义。大脑图像是医学图像分割最主要的研究领域之一。其中，大脑感兴趣区域分割和大脑肿瘤分割是研究最多的两个分支。大脑感兴趣区域分割是大脑疾病预测、观察大脑发育以及脑功能连接计算的关键技术，脑肿瘤分割是研究和诊断脑肿瘤的最重要手段。目前，多种技术可以实现大脑感兴趣区域和脑肿瘤的自动分割，而卷积神经网络由于能自发从输入图像中学习高等语义特征，实现从图像到分割结果端对端的映射，从而成为当前最主流的分割方法。但目前在大脑图像分割中依然存在诸多问题。例如大脑感兴趣区域分割时，相邻脑区交界处分割非常差；脑肿瘤分割时，小肿瘤容易被漏分割，肿瘤内部组织划分也容易出错。而且，目前提出的网络并没有从数据特点和任务角度对网络进行增强，卷积神经网络在脑图像分割领域仍有很大的提升空间。

本论文针对脑部医学磁共振图像的分割问题，提出以下几种卷积神经网络分割模型，实现大脑感兴趣区域的自动分割以及大脑肿瘤的自动检测。本文的主要内容和创新点如下：

1. 提出一种结合多图谱的分割网络对大脑感兴趣区域进行分割。本网络以3D图像块作为输入，将多个图谱信息作为目标图像块的附加信息融合到全卷积神经网络中。图谱中的标签信息对脑区边界处的分割有很强的指导作用。本方法融合了基于图谱分割和基于学习分割方法的优点，既避免了花费大量时间进行非刚性配准，又充分利用了图谱先验知识。实验表明，在LONI数据集的分割结果达到了80.3%的体积重合率。

2. 提出一种多通道多图谱引导的U-Net网络分割大脑感兴趣区域。该网络包含三个通道：图谱处理通路、目标图像处理通路和融合通路。每一个备选图谱，与目标图像拼接在一起共同输入到图谱处理通路中进行运算，通过网络不同层的卷积操作，使错误的图谱信息得到修正。另一方面，融合通路将多个独立的图谱通路进行加权融合，汇入目标通路中。融合通路采用多尺度融合策略，能确保充分利用不同尺度的图谱信息。目标通路对目标块以及融合后的图谱信息进行卷积操作，获得最终的分割结果。实验表明，在LONI数据集上的分割结果达到了81.2\%体积重叠率，在SATA数据集上的分割结果达到了89%的体积重叠率。

3. 提出一种三通道神经网络对大脑肿瘤进行分割。本网络利用不同模态影像显示肿瘤不同区域的特性对肿瘤进行分割。网络采用T1ce模态和FLAIR模态数据作为网络输入，T1ce模态图片通过单独的通路分割肿瘤的增强组织和坏死组织，FLAIR模态图片输入另一个单独的通路分割肿瘤整体。两个通路分别用不同的金标准信息进行监督学习。在网络的最后，融合通路对这两路输出特征进行合并，利用包含所有标签的金标准进行监督学习。同时，我们采用了一种添加权重的方法对网络损失函数进行优化，来减少样本不均衡对训练造成的影响。最后，利用后处理技术将分割结果中孤立的孤岛结果进行删除，获得更加精确的分割结果。

4. 提出了一种定位掩码分割网络分割大脑肿瘤。网络包含定位通路、掩码通路以及分割通路。特征提取网络生成的图像特征直接输入三个通路中。在定位通路，通过全连接映射获得包含肿瘤区域的两个角点的预测坐标。这两个角点坐标用于构建掩码图像，对输入掩码通路的图像特征进行过滤。这样，后续的卷积操作将只作用于区域内的特征。分割通路采用孔洞卷积策略，获得全局的特征信息。该图像特征与掩码通路生成的特征拼接在一起，共同预测整个肿瘤。本网络解决了常用的神经网络或者侧重于捕捉全局特征，或者侧重于聚焦于特定区域这个局限性。此外，还提出了一种图像增强的方法对训练数据进行处理，平衡由训练样本不均衡引起的训练模型对某些数据分割效果很差的问题。

Image segmentation is the most commonly used and important technique for medical image analysis. The brain is the most important organ in our body, the research on the brain is crucial for us better understanding our thoughts, understanding the development of disease and guiding the growth of our brains. Brain image segmentation has become the hottest topic in the medical image field. Brain ROI and brain tumor segmentation tasks are two main branches in brain segmentation area. Brain ROI segmentation is the key technique for predicting brain disease and calculating brain ROI connection, while brain tumor segmentation is the most important step for tumor diagnose. Currently, many automatic segmentation methods are proposed to segment the brain MRI. Convolutional neural networks based methods are the mainstream for this task since they could automatically learn semantic features from the input image. Moreover, they have achieved the-state-of-art in most medical image segmentation tasks. However, they still meet a lot of troubles when segmenting brain image. For instance, the segmentation along the ROI boundaries is very poor; small tumors are easily neglected; segmentations inside tumor are very bad. Moreover, the existing convolutional neural networks neither considered the trait of input image nor they are task-oriented designed, which lead to a bad performance in brain image segmentation.

This study proposed several automatic segmentation architectures to solve brain ROI (region of interest) segmentation and brain tumor segmentation problems from MRI. The main contributions and contents are listed below:

1.  Multi-Atlas Guided 3D Fully Convolutional Networks for Brain ROI segment. The 3D image patches were used as input to this network. Multi-atlase patches, which were used as the supplementary features, were also inputted into the proposed structure. The label maps in atlas contain strong semantic information, especially it contains clear edge information. This is a very useful guidance for ROI segmentation, especially for segmenting the boundaries of the ROI. This network combines the advantage of both registration-based and pathch-based methods, which can fully utilize the prior atlas information and can also aviod spend too much time in non-rigid registration. Experiments on LONI dataset show that the DSC can reach to 80.3%.

2.  Improve the Multi-Atlas Guided Network by Utilizing multi-pathways network. This network contains three pathways: atlas-unique pathway, atlas-aware fusion pathway, and target-patch pathway. The atlas-unique pathway can revise the wrong labels in the atlas by using the convolution operation. The atlas-aware fusion pathway gives each voxel in the candidate atlas patch a weight and fuses them together at the voxel level. A hierarchical strategy is applied to further fuse the atlas information. Last, the target-patch pathway propagates the target patch and the fused information. Experiments on LONI dataset and SATA dataset show that the DSC can reach to 81.2% and 89%, respectively.

3.  Three Pathways U-Net for Brain Tumor Segmentation. We utilize the fact that different modality images clearly show different sub-regions of the tumor. For our network, we use T1ce modality and FLAIR image as input. Each modality is processed in a single pathway, FLAIR path-way for whole tumor segmentation and t1ce pathway for the enhance tumor and necrotic components segmentation. At the end of the model, these two pathways are fused together by the fusion pathway to get the final segmentation. Extra losses are added in the model to guide the network training. The ground truth of the flair-pathway contains the healthy brain label and the whole tumor label. The ground truth of the t1ce-pathway contains the enhance tumor label and necrotic components label. The ground truth of the fusion-pathway contains all labels need to be segmented. Moreover, a weighting scheme at the loss layer is applied to alleviate the problem of label imbalance. At last, we delete the island segmentation to achieve a more precise result.

4.  Loc-Net for Brain Tumor Segmentation. The proposed architecture contains three pathways: Locating pathway, Mask pathway, and Segmentation pathway. Locating pathway predicts corner positions of the tumor bounding box. These corner positions are transformed to mask features generated by the residual network so that the later convolution operations can focus on the content inside the bounding box in the mask pathway. Segmentation pathway uses dilate convolution to capture the global context. These features are fused with masked features generated in the mask pathway to get the final prediction. Loc-Net can capture both the global context and simultaneously focus on the tumor region. Furthermore, we utilize a novel data augment strategy to augment the training data to improve the robustness of the model.