基于组织微阵列的乳腺癌自动诊断

发布时间：2018-12-19 10:01

【摘要】：在临床医院,基于乳腺癌组织病理图像评分的诺丁汉分级系统是由病理医生观察分析腺管的形成、核的异型性以及有丝分裂次数这三个指标得到的。由于经验和知识水平的差异,不同的医生对病理组织切片评分可能会有差异。因此病理组织分析的计算机辅助诊断研究具有重要意义,它能够为不同经验的临床医生提供客观的诊断结果,也可以避免一些人为的漏检。上皮和间质组织是乳腺组织中最为基本的两种组织,研究表明大约80%乳腺癌起源于上皮组织,因此上皮组织和间质组织及其微环境的分析是评估乳腺癌风险的重要标志。因此上皮组织和间质组织精确分割是构建计算机辅助诊断系统的前提条件。针对以上问题,本文研究了基于组织微阵列的乳腺癌自动诊断方法,该方法主要分为乳腺组织微阵列的上皮与间质组织的自动分割以及基于乳腺组织微阵列的癌症等级自动评分两个方面。具体内容为:首先采用基于全卷积网络的上皮与间质组织自动分割方法,实现端到端的小尺寸图像分割,从实验分割评估结果看出:在荷兰癌症研究所(NKI)和温哥华综合医院(VGH)提供的数据集上分割精确度(像素准确率92.0%、像素平均准确率85.7%、平均重叠率79.5%、类权重重叠率86.1%)最高,分割效果最好,同等条件下分割速度较快(0.097秒);然后通过滑动窗取块的方式批量输入网络实现大尺寸的组织微阵列图上皮与间质自动分割。最后提取出组织微阵列图像的上皮组织和间质组织后,分别进行提取特征(颜色直方图、纹理)并且组合特征,再把组合特征输入到支持向量机分类器中,得到病理分级结果(一级分类准确率81.7%、二级的分类准确率80.6%),针对实际的乳腺癌组织微阵列数据集实现乳腺癌组织病理自动分级,从而达到计算机辅助乳腺癌诊断的目标。
[Abstract]:In clinical hospitals, the Nottingham grading system based on the pathological image of breast cancer was obtained by the pathologist's observation and analysis of the formation of the glandular duct, the heterogeneity of the nucleus and the frequency of mitosis. Due to differences in experience and level of knowledge, different doctors may have different scores on histopathological sections. Therefore, the computer-aided diagnosis of pathological tissue analysis is of great significance. It can provide objective diagnostic results for clinicians with different experiences, and avoid some artificial misdiagnosis. Epithelium and mesenchymal tissue are the two most basic tissues in mammary gland. Studies show that about 80% of breast cancer originated from epithelial tissue, so the analysis of epithelium and interstitial tissue and its microenvironment is an important marker to assess the risk of breast cancer. Therefore, accurate segmentation of mesenchymal tissue and epithelial tissue is the precondition of computer aided diagnosis system. In view of the above problems, this paper studies the automatic diagnosis method of breast cancer based on tissue microarray. This method is mainly divided into two aspects: the automatic segmentation of epithelial and interstitial tissue of breast tissue microarray and the automatic grading of cancer grade based on breast tissue microarray. The main contents are as follows: firstly, an automatic segmentation method of epithelium and mesenchymal tissue based on full convolution network is adopted to realize end to end small size image segmentation. From the results of the experimental segmentation evaluation, we can see the accuracy of segmentation on the data set provided by the Dutch Cancer Institute (NKI) and the Vancouver General Hospital (VGH) (the pixel accuracy is 92.0, the average pixel accuracy is 85.775, The average overlap rate was 79.5%, the class weight overlap rate was 86.1%), the segmentation effect was the best, and the speed of segmentation was faster (0.097 seconds) under the same conditions; Then the large-scale tissue microarray epithelium and mesenchymal tissue are automatically segmented by bulk input network with sliding window. Finally, after extracting the epithelium and mesenchymal tissue of the tissue microarray image, the features (color histogram, texture) are extracted, and then the combined features are input into the support vector machine classifier. The results of pathological grading (81.7% of primary classification accuracy and 80.6% of second-level classification accuracy) were obtained. According to the actual breast cancer tissue microarray data set, the automatic classification of breast cancer tissue was realized. In order to achieve the goal of computer-aided breast cancer diagnosis.
【学位授予单位】：南京信息工程大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R737.9;TP391.41

【参考文献】