Classification On Esophageal Cancer X-ray Image Of Xinjiang Kazak Based On Data Mining

Objective:The aim of this research is to study application of data mining in esophageal cancer X-ray image classification in Xinjiang Kazakh so as to provide valuable assistant information for radiologists in diagnosing and thus improve diagnosing precision and efficiency.Methods:The experiments were conducted in the MATLAB platform.Firstly,the regions of interest were selected by hand.And the preprocessing methods,including median filter and histogram equalization,were applied on the X-ray images.This step can improve the quality of the images.Secondly,four algorithms,including gray-level histogram,gray level co-occurrence matrix,gray level gradient co-occurrence matrix and Tamura texture were employed to the image feature extraction.Thirdly,PCA feature selection method and AUC area constraint feature screening method were used to optimize the extracted features,and the redundant feature quantity was eliminated.By using KNN,RF,SVM and Logistic regression classification,the esophageal X-ray images of normal esophageal and early esophageal cancer,ulcerative type,constrictive type and fungating type were evaluated by 10-folded cross validation.Various model parameters were evaluated by ROC and Calibration curve.Results:Normal esophagus and early esophageal cancer types:a)PCA feature selection:extracted to eight principal components,in the PCA feature set KNN(K = 1),RF,SVM and Logistic regression classifier on normal esophageal classification accuracy rate of 91.43%,80.38%,89.26%and 95.29%respectively.The accuracy rate of early esophageal cancer was 94.29%,84.67%,92.65%and 97.14%respectively.b)AUC area constraint feature screening:18 AUC values greater than 0.75 were selected.The accuracy of KNN(K = 3),RF,SVM and Logistic regression classifier on normal esophageal classification was 88.60%,78.57%,92.96%and 86.37%respectively.The accuracy rate of early esophageal cancer was 86.50%,82.65%,94.29%and 85.86%respectively.Ulcerative,constrictive and fungating esophagus cancers:a)PCA characteristics selection:selecting 7 main components.The accuracy rate of KNN(K= 1),RF,SVM and Logistic regression classifier on ulcerative esophagus cancer was 88.13%,87.32%,93.67%and 94.56%respectively.The accuracy rate on constrictive esophageal cancer was 87.90%,88.64%,92.67%and 94.30%respectively.And the accuracy rate on fungating esophagus cancer was 86.88%,85.12%,91.54%and 93.36%respectively.b)AUC area constraint feature screening:14 characteristics quantities were selected with all their AUC values exceeding 0.75,and the accuracy rate of KNN(K = 3),RF,SVM and Logistic regression classifier for ulcerative esophageal cancer was 83.45%,87.34%,94.45%and 95.68%respectively.The accuracy rate on constrictive esophageal cancer was 84.75%,83.36%,91.33%and 94.05%respectively.And the accuracy rate fungating esophagus cancer was 83.55%,79.42%,90.58%and 91.36%respectively.Conclusion:The normal esophagus and early esophageal carcinoma classification:PCA feature selection method is superior to AUC area constrained feature screening method.Early esophageal cancer classification is better than normal esophagus.While,SVM classifier classification is the best.Ulcerative type,constrictive type,and fungating type:PCA feature selection method and AUC area constraint feature screening method are more suitable for advanced esophageal cancer;Ulcer type esophageal cancer classification is the best;Logistic regression and SVM classifier are more suitable for the classification of advanced esophageal cancer.The classification results between early and advanced esophageal cancer,PCA feature selection and SVM classifier were most suitable.The results of this study can provide valuable reference for radiologists to diagnose esophageal cancer,especially for the diagnosis of early esophageal cancer,which lays the foundation for the development of computer-aided diagnosis system for Xinjiang Kazakh esophageal cancer.