基于数据挖掘技术的新疆哈萨克族食管癌X线图像的分型研究

发布时间：2018-01-31 23:45

本文关键词： 食管癌图像分型数据挖掘技术计算机辅助诊断分类模型评估　出处：《新疆医科大学》2017年硕士论文　论文类型：学位论文

【摘要】：目的:本研究针对食管X线图像,研究数据挖掘技术在新疆哈萨克族食管癌X线图像分型中的应用,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、灰度梯度共生矩阵和Tamura纹理特征;采用PCA特征选择法和AUC面积约束特征筛选法对所提取到的特征进行优化选择,剔除冗余特征量;使用KNN、RF、SVM和Logistic回归分类器,通过10折交叉验证,对正常食管和早期食管癌,溃疡型、缩窄型、蕈伞型中晚期食管癌食管X线图像进行分类研究;利用参数评估、ROC和Calibration曲线对各分类模型性能进行评价。结果:正常食管和早期食管癌分型:a)PCA特征选择:提取到8个主成分,在PCA特征集下KNN(K=1)、RF、SVM和Logistic回归分类器对正常食管分类准确率为91.43%、80.38%、89.26%和95.29%,对早期食管癌分类准确率为94.29%、84.67%、92.65%和97.14%。b)AUC面积约束特征筛选:筛选到18个AUC值大于0.75的特征量,在AUC特征集下KNN(K=3)、RF、SVM和Logistic回归分类器对正常食管分类准确率88.60%、78.57%、92.96%和86.37%,对早期食管癌分类准确率为86.50%、82.65%、94.29%和85.86%。溃疡型、缩窄型、蕈伞型食管癌分型:a)PCA特征选择:提取到7个主成分,在PCA特征集下KNN(K=1)、RF、SVM和Logistic回归分类器对溃疡型食管癌分类准确率为88.13%、87.32%、93.67%和94.56%,对缩窄型食管癌分类准确率为87.90%、88.64%、92.67%和94.30%,对蕈伞型食管癌分类准确率为86.88%、85.12%、91.54%和93.36%。b)AUC面积约束特征筛选:筛选到14个AUC值大于0.75的特征量,在AUC特征集下KNN(K=3)、RF、SVM和Logistic回归分类器对溃疡型食管癌分类准确率为83.45%、87.34%、94.45%和95.68%,对缩窄型食管癌分类准确率为84.75%、83.36%、91.33%和94.05%,对蕈伞型食管癌分类准确率为83.55%、79.42%、90.58%和91.36%。结论:正常食管和早期食管癌分型:PCA特征选择法优于AUC面积约束特征筛选法;早期食管癌分类效果优于正常食管;SVM分类器分类效果最佳。溃疡型、缩窄型、蕈伞型食管癌分型:PCA特征选择法和AUC面积约束特征筛选法均适用于中晚期食管癌;溃疡型食管癌分类效果最好;Logistic回归和SVM分类器更适合于中晚期食管癌的分型;综合早期和中晚期食管癌分型结果,PCA特征选择和SVM分类器最适用于食管癌分型。本研究结果能为放射科医生对食管癌的诊断提供有价值的参考意见,尤其是对早期食管癌的诊断,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 application of data mining technology in the classification of esophageal carcinoma in Xinjiang Kazak nationality. The purpose of this study was to provide assistant information with practical reference value for radiologists to make diagnosis decision, and to improve the accuracy and efficiency of diagnosis of esophageal cancer in Kazak nationality in Xinjiang. Methods: MATLAB image processing software was used. In order to improve the image quality, the region of interest of esophageal X-ray image is manually selected, and the median filter denoising and histogram equalization are used to improve the image quality. Extraction based on gray histogram, gray co-occurrence matrix, gray-level gradient co-occurrence matrix and Tamura texture features; The PCA feature selection method and the AUC area-constrained feature selection method are used to optimize the selection of the extracted features, and the redundant feature quantities are eliminated. By using Logistic regression classifier and Logistic regression classifier, normal esophagus, early esophageal carcinoma, ulcerative type and constriction type were examined by 10 fold cross validation. The esophageal X-ray images of middle and late stage esophageal carcinoma of mushroom type were studied. Using parameter evaluation. ROC and Calibration curves were used to evaluate the performance of each classification model. Results: eight principal components were extracted from normal esophagus and early esophageal carcinoma. Under the PCA feature set, the accuracy of PCA and Logistic regression classifier for normal esophagus classification was 91.43% and 80.38% respectively. The accuracy of classification of early esophageal carcinoma in 89.26% and 95.29 was 94. 29% and 84. 67%. 92.65% and 97.14. The area constraint characteristics of AUC were screened: 18 characteristic quantities with AUC value greater than 0.75 were screened, and the RF was obtained under the AUC feature set. The accuracy of SVM and Logistic regression classifier for normal esophageal classification was 88.60% and 92.96% and 86.37% respectively. The accuracy of classification for early esophageal carcinoma was 86.50%, 82.65% and 85.86%, respectively. The selection of PCA features for the classification of mushroom esophageal carcinoma: seven principal components were extracted, and the RF of KNNN was extracted under the PCA feature set. The accuracy of SVM and Logistic regression classifier in the classification of ulcerative esophageal carcinoma was 88.13 ~ 87.32 ~ 93.67% and 94.56% respectively. The accuracy of classification of constricted esophageal carcinoma was 87.90% and 94.30%, 86.8888% and 85.12% respectively. 91.54% and 93.36. area restriction feature screening of AUC: 14 characteristic quantities with AUC value greater than 0.75 were screened, and the RF of KNNs was found to be higher than 0.75 under the AUC feature set. The accuracy of SVM and Logistic regression classifier in the classification of ulcerative esophageal carcinoma was 83.45% and 95.68% respectively. The accuracy of classification of constricted esophageal carcinoma was 91.33% and 94.05%, and that of mushroom carcinoma was 83.55% and 79.42%, respectively. Conclusion: the classification of normal esophagus and early esophageal carcinoma is superior to that of AUC. The classification effect of early esophageal carcinoma was better than that of normal esophagus. The results of SVM classifier were the best. The ulcerative type, constriction type, mushroom type esophageal carcinoma typing: PCA: 1. The AUC area constraint feature selection method were all suitable for the middle and late stage esophageal carcinoma. The classification effect of ulcerative esophageal carcinoma was the best. Logistic regression and SVM classifier were more suitable for classification of advanced esophageal carcinoma. Combined with the early and late esophageal cancer classification results, the SVM classifier and PCA feature selection are the most suitable for the classification of esophageal cancer. The results of this study can provide valuable reference for radiologists in the diagnosis of esophageal cancer. Especially, the diagnosis of early esophageal cancer lays a foundation for the development of the computer aided diagnosis system for Xinjiang Kazak esophageal carcinoma.
【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

【相似文献】