新疆哈萨克族食管癌X射线图像计算机辅助诊断技术的研究

发布时间：2018-08-02 09:01

【摘要】：目的:本研究针对食管X射线图像,研究适用于新疆哈萨克族食管癌计算机辅助诊断中涉及到的技术和方法,旨在为放射科医生的诊断决策提供具有实际参考价值的辅助信息,提高新疆哈萨克族食管癌诊断的准确率和效率。方法:利用MATLAB图像处理软件,手动选取食管X射线图像的感兴趣区域,对其进行中值滤波去噪和直方图均衡化的预处理以改善图像质量;提取基于灰度直方图、灰度共生矩阵、灰度梯度共生矩阵、Tamura纹理、小波频域和Hu不变矩特征;提出采用ROC曲线下面积选择法和主成分分析相结合的两步式特征选择法对所提取的特征进行优化选择,消除各种单一特征间的冗余信息;使用Lib-SVM工具箱,通过10折交叉验证,选取基于RBF核函数的SVM,对正常食管和病理食管图像,蕈伞型、浸润型和溃疡型食管X射线图像进行分类研究,采用分类准确率和ROC曲线下面积AUC值作为分类器性能的评价指标。结果:研究共计提取66维特征,根据ROC曲线下面积AUC值大于0.7的原则筛选出28维特征,在此基础上进行主成分分析,选择前10个累计贡献率达到90.33%的主成分。利用单一特征进行分类时,灰度共生矩阵特征表现出最佳的分类性能,对正常食管、病理食管图像的分类准确率和AUC值分别为92.67%和92.70%,分类运行时间为2.40s;对蕈伞型、浸润型和溃疡型食管癌图像的分类准确率和AUC值分别为90.67%和91.00%,分类运行时间为2.80s;选用综合特征进行分类时,对正常食管、病理食管图像的分类准确率和AUC值均为88.50%,分类运行时间为7.70s;对蕈伞型、浸润型和溃疡型食管癌图像的分类准确率和AUC值分别为88.40%和88.53%,分类运行时间为9.70s;采用AUC值大于0.7的特征进行分类时,对正常食管、病理食管图像的分类准确率和AUC值分别为94.17%和94.20%,分类运行时间为2.80s;对蕈伞型、浸润型和溃疡型食管癌图像的分类准确率和AUC值分别为92.33%和92.30%,分类运行时间为4.50s;利用主成分分析筛选的特征进行分类时,对正常食管、病理食管图像的分类准确率和AUC值分别为95.33%和95.00%,分类运行时间为1.50s;对蕈伞型、浸润型和溃疡型食管癌图像的分类准确率和AUC值分别为94.59%和94.00%,分类运行时间为2.30s。结论:本研究对正常食管、蕈伞型、浸润型和溃疡型食管癌X射线图像进行特征提取、特征选择及图像分类,采用将两步式特征选择与SVM相结合的方法,取得了较高的分类准确率,能为放射科医生提供有价值的参考意见,辅助其更好地诊断食管癌,为开发面向临床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 techniques and methods involved in computer-aided diagnosis of esophageal carcinoma in Kazak nationality in Xinjiang, in order to provide practical reference information for radiologists to make diagnosis decisions. To improve the accuracy and efficiency of diagnosis of esophageal cancer in Kazak nationality in Xinjiang. Methods: using MATLAB image processing software, the region of interest of esophageal X-ray image was manually selected, and the median filter denoising and histogram equalization were performed to improve the image quality, and the gray level co-occurrence matrix was extracted based on gray histogram. The grayscale gradient co-occurrence matrix is Tamura texture, wavelet frequency domain and Hu moment invariant feature, and a two-step feature selection method combining area selection method under ROC curve and principal component analysis is proposed to optimize the feature selection. Using Lib-SVM toolbox, 10 fold cross validation, select SVM based on RBF kernel function, for normal esophagus and pathological esophagus image, mushroom type, The classification of invasive and ulcerative esophageal X-ray images was studied. The accuracy of classification and the AUC value of area under the ROC curve were used to evaluate the performance of the classifier. Results: a total of 66 dimensional features were extracted and 28 dimensional features were selected according to the principle of AUC value greater than 0. 7 under the ROC curve. On this basis principal component analysis was carried out and the first 10 principal components with cumulative contribution rate of 90.33% were selected. For normal esophagus and pathological esophagus, the classification accuracy and AUC value were 92.67% and 92.70%, respectively. The classification time was 2.40 s. The classification accuracy and AUC value of invasive and ulcerative esophageal carcinoma images were 90.67% and 91.00%, respectively, and the classification time was 2.80 s. The classification accuracy and AUC value of pathological esophageal images were 88.50 and 7.70 s respectively. The classification accuracy and AUC value of invasive and ulcerative esophageal carcinoma images were 88.40% and 88.53, respectively. The classification time was 9.70 s. When the AUC value was greater than 0.7, the normal esophagus was classified. The classification accuracy and AUC value of pathological esophageal images were 94.17% and 94.20%, respectively, and the classification time was 2.80 s. The classification accuracy and AUC value of invasive and ulcerative esophageal carcinoma images were 92.33% and 92.30%, respectively, and the operating time of classification was 4.50 s. The classification accuracy and AUC value of pathological esophageal images were 95.33% and 95.00%, respectively, and the classification time was 1.50 s, and the classification accuracy and AUC value of fungote, infiltrative and ulcerative esophageal carcinoma images were 94.59% and 94.00%, respectively, and the running time of classification was 2.30 s. Conclusion: in this study, the X-ray images of normal esophagus, mushroom type, infiltrating type and ulcerative type of esophageal carcinoma were extracted, selected and classified, and the method of combining two-step feature selection with SVM was used. It can provide valuable reference for radiologists and help them to diagnose esophageal cancer better. It lays a foundation for the development of the computer aided diagnosis system for Xinjiang Kazak nationality esophageal cancer.
【学位授予单位】：新疆医科大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R735.1

【相似文献】