基于计算机辅助识别新疆高发病食管癌图像的算法研究

发布时间：2018-06-01 20:39

本文选题：新疆哈萨克族 + 食管癌　；参考：《新疆医科大学》2017年硕士论文

【摘要】：目的:本研究根据新疆哈萨克族食管X射线图像的特点,探讨计算机辅助诊断技术和算法在新疆哈萨克族食管癌图像中的应用,验证算法的可行性及合理性。从而为基于医学X射线图像的计算机辅助诊断技术辅助医生进行病变类型图像的判别,降低医生的工作量,提高诊断质量。方法:利用MATLAB图像处理软件,对食管X射线图像进行预处理包括选取感兴趣区域、中值滤波和直方图均衡化;其次,将预处理后的图像选用阈值分割法进行分割;利用灰度共生矩阵、Hu不变矩特征、灰度直方图和小波变换算法对分割后的图像提取特征值;采用两种特征选择算法,基于独立样本T检验和主成分分析相结合对正常食管和病理食管图像的特征值进行优化选择,基于单因素方差分析和主成分分析结合的特征选择算法对蕈伞型、溃疡型和浸润型食管癌图像进行特征筛选,消除不同特征间的冗余信息;根据上述两种特征选择算法,采用BP神经网络分别设计两个分类器对哈萨克族食管X射线图像进行分类识别,采用分类准确率和Kappa值作为分类器性能的评价指标。结果:(1)经过预处理后图像的质量得到了提高,图像的清晰度明显增强,且获得的图像边缘细节都较清晰;(2)本文使用阈值分割法对预处理后的图像进行分割,可以分割出完整的、清晰的、不失真的病灶区域,该方法保留了图像的完整性;(3)通过四种特征提取算法对分割后的正常图像和病理图像进行特征的提取,共计提取了29个特征值;(4)第一种特征选择算法,即使用独立样本T检验结合主成分分析特征选择算法对正常食管和病理食管进行特征筛选时,首先使用独立样本T检验筛选出19个特征值,然后在此基础上对特征值做主成分分析,选择前5个累积贡献率达到88.085%的主成分;另一种方法为采用单因素方差分析和主成分分析结合的特征选择算法对蕈伞型、溃疡型和浸润型食管癌图像进行特征的筛选时,首先利用单因素方差分析选择了24个特征,然后使用主成分分析对这些特征值进行计算,所以求得主成分的个数为5,且累积贡献率达到了87.537%;(5)使用BP神经网络对正常食管和病理食管进行分类时,独立样本T检验结合主成分分析特征选择算法在隐含层节点数为7时对图像的分类结果高于其它四种不同特征提取算法、综合特征及独立样本T检验筛选出的特征值的分类结果,即正常和病理食管图像的平均分类准确分别为97.130%、98.620%,分类效果较好;使用BP神经网络结合单因素方差分析与主成分分析特征选择算法对三种食管癌图像进行分类时,所得蕈伞型、溃疡型和浸润型食管癌的分类准确率分别为98.000%、96.000%、98.500%,与其它三种方法比较,分类结果较为精确;结论:本研究选取的预处理和阈值分割算法不仅去除了噪声使图像质量得到改善,而且也保留了完整的目标区域。其次,对处理后的图像使用不同特征提取算法得到的特征值采用特征选择算法筛选出了分类能力较强的特征值。最后,利用BP神经网络分类器对食管图像进行分类,且取得了较高的分类准确率。本文提出的特征选择算法结合BP神经网络算法是合理可行的,这为哈萨克族地区的放射科医生提供有价值的参考意见,提高诊断质量,为开发面向放射科的新疆哈萨克族食管癌计算机辅助诊断系统奠定了基础。
[Abstract]:Objective: To study the application of computer aided diagnosis technology and algorithm in the Xinjiang Kazak esophageal cancer image in Xinjiang based on the characteristics of the Kazak's X ray image of the Kazak nationality in Xinjiang, and to verify the feasibility and rationality of the algorithm. Thus, the computer aided diagnosis technology based on the medical X ray image is used to assist the doctor to carry out the disease type image To reduce the workload of the doctors and improve the quality of diagnosis. Methods: using MATLAB image processing software, the preprocessing of the X ray images of the esophagus includes selected regions of interest, median filtering and histogram equalization; secondly, the pre processed images are segmented by threshold segmentation, and the gray level symbiotic matrix and Hu invariant moments are used. The gray histogram and wavelet transform algorithm extracts the eigenvalues of the segmented images. Two feature selection algorithms are used to select the eigenvalues of the normal esophagus and the pathological esophagus, based on the combination of independent sample T test and principal component analysis, and the feature selection algorithm based on the combination of single factor variance analysis and principal component analysis. The images of fungoid, ulcerative and infiltrating esophageal cancer were screened to eliminate the redundant information between different features. According to the above two feature selection algorithms, two classifiers were designed by BP neural network to classify the Kazak's esophagus X ray images, and the classification accuracy and Kappa values were used as the classifier performance evaluation. Results: (1) the quality of the image is improved after preprocessing, the image sharpness is obviously enhanced and the image edge details are clear. (2) this paper uses the threshold segmentation method to segment the pre processed image, which can separate the complete, clear and undistorted focus area, which preserves the image. Integrity; (3) four feature extraction algorithms are used to extract the characteristics of the normal and pathological images after the segmentation, and a total of 29 eigenvalues are extracted. (4) the first feature selection algorithm, the first use of the independent sample T test combined with the principal component analysis feature selection algorithm to screen the normal esophagus and the pathological esophagus. 19 eigenvalues were screened by independent sample T test, then the principal component analysis was performed on the eigenvalues, and the first 5 cumulative contribution rates of 88.085% were selected, and the other was a feature selection algorithm combined with single factor analysis of variance and principal component analysis for the image of fungoid, ulcerative and infiltrating esophageal cancer. When screening, 24 characteristics are selected by single factor analysis of variance, and then the principal component analysis is used to calculate these eigenvalues, so the number of the principal components is 5 and the cumulative contribution rate is 87.537%. (5) the independent sample T test combined with the main formation using the BP neural network for the normal esophagus and the pathological esophagus. The classification result is higher than the other four different feature extraction algorithms when the number of hidden layer nodes is 7. The classification results of the characteristic values selected by the integrated feature and the independent sample T test, that is, the average classification accuracy of the normal and pathological esophagus images is 97.130%, 98.620%, and the classification effect is better; the use of BP God is good. The classification accuracy of the three kinds of esophageal cancer was 98%, 96%, 98.500%, respectively, with the single factor variance analysis and the principal component analysis feature selection algorithm. The classification results were more accurate than those of the other three methods. Conclusion: This study selected the preprocessing and the results of the study. The threshold segmentation algorithm not only removes the noise to improve the quality of the image, but also preserves the complete target area. Secondly, the eigenvalues obtained by using different feature extraction algorithms after the processing of the processed images are selected by the feature selection algorithm. Finally, the BP neural network classifier is used for the esophagus map. It is reasonable and feasible to combine the feature selection algorithm combined with the BP neural network algorithm. This provides a valuable reference for the radiologists of the Kazak region to improve the quality of diagnosis and to develop a computer aided diagnosis of the Xinjiang Kazakh cancer in the radiology department. The system lays the foundation.
【学位授予单位】：新疆医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R735.1

【参考文献】