CASIA OpenIR  > 博士后  > 出站报告
基于深度多层表示的乳腺癌病理图像细胞有丝分裂检测方法
杨晨雪
2018-07
关键词深度学习 有丝分裂细胞检测 生成对抗网络 多粒度级联森林方法 框回归过程 低秩表示
中文摘要
Cancer is one of the major diseases with the highest incidence, which imposes serious threats to human health. Pathological cell image analysis is a key step in the diagnosis of cancer tumors, and is also recognized as the "gold standard". The counting of number of mitotic cells in histopathological images is one of the important indicators of cancer grading assessment. Due to the nature of the biological cell image, ordinary people cannot identify whether the cell is undergoing mitosis in the cancer tissue cell image, so a large number of experts are required to label the mitotic cell sample. However, the process of manually detecting and calculating mitotic cells is very time consuming, and the diagnosis results vary greatly among doctors. The severe deficiency of mitotic cell data has caused a non-negligible impact on the accuracy of automated detection of mitotic cells in computer-aided cancer pathology images. Therefore, there is an urgent need in the clinic for methods of automated detection and enumeration of mitotic cells.
In the second chapter, we first introduce several different forms of generative adversarial networks (GANs), and discuss in detail the research progress and application of these GANs in modeling, architecture, training and performance evaluation. Then, based on the appearance changes of the cell morphology of cancer cells, normal cells and mitotic cells, the CycleGAN was selected as a base method. Through unsupervised learning combined with loss of resistance and loss of circulation consistency, the mapping relationship between mitotic cell images and normal cell images, mitotic cell images, and cancerous cell images in breast cancer H&E-stained pathological images was obtained, and different cells were realized. The mutual conversion between type images generates more mitotic cell images. After verification by pathologists, these generated mitotic cell samples have high authenticity.
In Chapter 3, we present a region detection method based on multi-grainularity cascading forests (Regions with multi-Grained Cascade forest features, RgcForest), which applies the gcForest classification method to the cell mitosis detection task. First, the sliding window is added to the original algorithm, and then the image regions is fed and the results of “is mitosis cell or not” are produced. If it is a mitosis cell, the coordinates are also provided. Then, in order to reduce the unnecessary bounding box in the detection result, a non-maximal suppression algorithm is implemented on the gcForest algorithm. In the end, similar as the evolution from CNN to RCNN, the process of bounding box regression is implemented to further improve the localizatoin accuracy. The data generated by CycleGAN in the second chapter is used as the positive samples, and selects cancerous cells, normal cells, and cytoplasmic background images as negative samples. We train our gcForest-based method, and then uses the detection framework to effectively and accurately detection of mitotic cells.
In chapter 4, a mitotic detection method for breast cancer pathology image is proposed which is based on cascade low-rank multi-level decomposition representation. The non-mitotic part of different slices in the same patient is considered as a low-rank part in the matrix representations. Since mitotic cells have different shapes at different periods, and different nucleus in the same period have different shapes and textures during mitosis, they are regarded as sparse parts in the low rank decomposition. We adopt an iterative way to perform low-rank decomposition on the original data matrix, which not only reduces the rank of the new low-rank matrix gradually, but also avoid the problem of image representation capability degradation caused by unreasonable parameter selection. The model can be learned layer-by-layer. The number of Mitochondria candidates are removed as much as possible in the layer-by-layer matrix low-rank decomposition. For different levels of matrix decomposition representation, the hierarchical classifier corresponding to each layer is designed and trained. All the image patches containing mitotic cells and non-mitotic parts decomposed from each representation layer are fed into the corresponding layer-specific classifiers, and the mitotic cells in the pathological image of breast cancer can be accurately detected. Compared with existing methods, the proposed method can quickly and conveniently obtain the detection results, and does not require cell segmentation and additional model learning, providing a possibility for clinical application.
 
英文摘要
癌症是发病率和死亡率最高的重大疾病之一,严重威胁着人类健康。病理细胞图像分析是癌症肿瘤诊断中关键的步骤,也是癌症肿瘤诊断的“黄金标准”。组织病理图像中有丝分裂细胞的计数是癌症分级评估的重要指标之一。由于生物细胞图像本身属性,普通人无法在癌症切片组织细胞图像中识别细胞是否在进行有丝分裂,故需要大量的专家来对有丝分裂细胞样本进行标注。然而,人工检测和计算有丝分裂细胞的过程非常冗长,而且不同医生之间的诊断结果有较大差异性。有丝分裂细胞数据的严重匮乏,对计算机辅助癌症病理图像的有丝分裂细胞的自动检测准确率造成了不可忽视的影响。因此,在临床中迫切需要能够定量地通过计算机实现癌症病理图像的有丝分裂细胞的自动检测和计数的方法。
本文第二章介绍了几种不同形式的生成式对抗网络,详细论述这几种生成式对抗网络在建模、架构、训练和性能评估方面的研究进展及其具体应用现状。然后,根据癌细胞、正常细胞和有丝分裂细胞的细胞形态的外观变化,选择循环一致对抗神经网络(CycleGAN)模型作为研究工具。通过将对抗损失与循环一致性损失相结合的无监督学习,分别得到乳腺癌H&E染色病理图像中的有丝分裂细胞图像和正常细胞图像、有丝分裂细胞图像和癌变细胞图像之间的映射关系,实现不同细胞类型图像之间的相互转换,生成更多的有丝分裂细胞图像。经过病理医师的验证,这些生成的有丝分裂细胞样本具有较高真实性。
本文第三章提出了一个基于多粒度级联森林的特征区域检测方法,(Regions with multi-Grained Cascade forest features,RgcForest),将gcForest分类方法应用到乳腺肿瘤细胞图像的细胞有丝分裂检测问题,借助gcForest良好的特征提取和分类性能,通过候选区域方法(Region Proposal)实现目标检测问题的转化。首先在原始算法上增加滑动窗口功能,然后把剪切好的图像给输入进去,并判断是否有丝分裂细胞,如果是则给出坐标。然后,为了减少检测结果中不必要的边框(bounding box),在gcForest算法上实现了非极大值抑制算法。最后,仿照CNN到RCNN的演进,增加边框坐标回归(bounding box regression)过程,进一步提高定位效果。实验选择第二章用CycleGAN模型生成的有丝分裂样本作为正样本,选择癌变细胞、正常细胞和细胞质背景图像作为负样本,在原始的gcForest框架上训练输出图像的概率值,再利用检测框架有效准确地检测有丝分裂细胞。
本文第四章提出了一种基于矩阵低秩多层分解表示级联分类器的乳腺癌病理学图像有丝分裂检测方法,将同一病患不同切片中的非有丝分裂部分看作是低秩表示模型中的低秩部分。由于有丝分裂的细胞在各个时期的形态各异,且同一时期不同的细胞核在进行有丝分裂阶段的形状和纹理变化也不尽相同,故将其看作是低秩表示中的稀疏部分。该方法采用迭代的方式对原始数据矩阵进行低秩分解,不仅可以使新的低秩矩阵的秩逐渐降低,避免了由于参数选择的不合理造成的图像表示能力下降的问题,而且在逐层的矩阵低秩多层分解中尽可能多地排除掉有丝分裂候选结果。针对矩阵低秩多层分解模型的不同层表示,设计和训练该层对应的级联分类器。将矩阵深度低秩分解模型中每一层分解出来的含有有丝分裂的细胞和非有丝分裂部分的所有图像块,分别输入对应的级联分类器中,可以准确地检测乳腺癌病理图像中的有丝分裂细胞。和己有的方法相比,所提方法能够快速方便的得到检测结果,不需要进行非常困难的细胞分割、模型训练等过程,为临床应用提供了一种可能。
文献类型研究报告
条目标识符http://ir.ia.ac.cn/handle/173211/21602
专题博士后_出站报告
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
杨晨雪. 基于深度多层表示的乳腺癌病理图像细胞有丝分裂检测方法. 2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
博士后出站报告-杨晨雪-最终版本.pdf(3442KB)研究报告 开放获取CC BY-NC-SA浏览
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[杨晨雪]的文章
百度学术
百度学术中相似的文章
[杨晨雪]的文章
必应学术
必应学术中相似的文章
[杨晨雪]的文章
相关权益政策
暂无数据
收藏/分享
文件名: 博士后出站报告-杨晨雪-最终版本.pdf
格式: Adobe PDF
此文件暂不支持浏览
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。