基于病理图像的乳腺肿瘤定量化分析

发布时间：2018-01-12 03:07

本文关键词：基于病理图像的乳腺肿瘤定量化分析　出处：《南京信息工程大学》2016年硕士论文　论文类型：学位论文

【摘要】：在乳腺癌诊断与预后过程中,通常由医生通过显微镜观察组织切片中不同的病理标志物对病理等级进行评分。然而人工分析的方式耗时且带有较强的医生主观性,不同医生的诊断结果存在不一致性,这可能会给患者带来严重的“过度治疗”和“治疗不当”。因此,研究计算机辅助诊断系统能为临床医生提供准确、定量的辅助分析结果从而加快治疗进程,对于医生和患者都具有重要意义。诺丁汉分级系统与细胞的外观和空间分布特征存在密切的联系,因此病理图像中细胞的检测与分割是构建病理图像自动分析系统的基础。然而,病理图像组织结构复杂成分众多,且细胞外观呈现高度的异质性。此外细胞之间存在重叠、挤压现象,因此细胞的检测与分割是一项极具挑战性的工作。针对这一问题,本文提出了基于深度卷积神经网络初始化的主动轮廓自适应椭圆拟合细胞分割方法。该方法运用卷积神经网络结合滑动窗口自动检测细胞,根据细胞检测结果初始化主动轮廓模型,最后使用自适应椭圆拟合方法分割重叠的细胞。为了验证该方法的性能,本文分别在三个数据集上进行了测试。实验结果表明：本文方法在三个数据集上的检测准确率分别为：73.33%,83.91%和76.88%,在数据集1和2上分割准确率分别为：85.03%,90.33%,说明本文方法性能优于其他对比方法。在自动病理分级研究中,本文提出了一个基于多特征描述的乳腺肿瘤病理自动分级方法。该方法使用卷积神经网络模型检测病理图像中的上皮细胞和淋巴细胞；然后运用颜色分离算法把细胞通道从HE染色的病理图像中分离,接下来使用自适应阈值、形态学操作、带有前景标记的分水岭算法和椭圆拟合得到细胞的边界。随后提取出细胞的形状纹理和反映细胞分布的空间结构特征,将这些特征降维后输入到支持向量机中实现对病理图像自动分级。实验结果表明：本文方法整体分类准确率为90.20%,对高中低各等级的区分准确率分别为92.87%,82.88%和93.61%,其性能远高于其他对比方法。
[Abstract]:In the diagnosis and prognosis of breast cancer. The pathological grade is usually scored by the doctor by observing the different pathological markers in the tissue slice by microscope. However, the manual analysis method is time-consuming and has a strong subjectivity of the doctor. The results of different doctors' diagnosis are inconsistent, which may bring serious "overtreatment" and "improper treatment" to patients. Therefore, the study of computer-aided diagnosis system can provide clinicians with accuracy. Quantitative analysis of the results to speed up the treatment process is of great significance to both doctors and patients. Nottingham grading system is closely related to the appearance and spatial distribution of cells. Therefore, the detection and segmentation of cells in pathological images is the basis of constructing an automatic analysis system for pathological images. However, there are many complex components in pathological images. Moreover, the appearance of cells is highly heterogeneous. In addition, there is overlap and squeezing between cells, so the detection and segmentation of cells is a very challenging task. In this paper, an active contour adaptive ellipse fitting cell segmentation method based on deep convolution neural network initialization is proposed, which uses convolution neural network combined with sliding window to automatically detect cells. The active contour model was initialized according to the results of cell detection, and then the overlapping cells were segmented by adaptive ellipse fitting method to verify the performance of the method. The experimental results show that the accuracy of this method on the three datasets is 83.91% and 76.88%, respectively. The segmentation accuracy on data set 1 and 2 is: 85.03 / 90.33, respectively, which shows that the performance of this method is superior to that of other comparison methods. In this paper, an automatic classification method for breast tumors based on multi-feature description is proposed, which uses convolution neural network model to detect epithelial cells and lymphocytes in pathological images. Then the color separation algorithm is used to separate the cell channel from the pathological image stained by HE. Then the adaptive threshold and morphological operation are used. The boundary of cells was obtained by watershed algorithm with foreground marker and ellipse fitting. Then the shape and texture of cells and the spatial structure characteristics reflecting the distribution of cells were extracted. These features are reduced and then input into support vector machine to realize the automatic classification of pathological images. The experimental results show that the overall classification accuracy of this method is 90.20%. The accuracy of distinguishing between high and low grades is 92.87% and 93.61%, respectively, and its performance is much higher than that of other comparison methods.
【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】：2016
【分类号】：TP391.41;R737.9

【相似文献】