RF和C4.5决策树在食管癌图像分类中的研究

目的 探讨RF和C4.5决策树对X线食管造影图像分型中的应用，以及验证分类器对特征的分类能力。方法 选取2018年1月~6月在新疆医科大学第一附属医院、第二附属医院和第三附属（肿瘤）医院的放射科选取溃疡性、缩窄型和蕈伞型食管癌X线图像各560张，提取灰度共生矩阵，灰度直方图和混合特征；采用RF和C4.5决策树通过调整参数进行分类研究。结果 RF和C4.5决策树对溃疡型和缩窄型食管癌进行分类，灰度共生矩阵的分类准确率分别为73.30%，67.76%；灰度直方图分类准确率分别为84.55%，76.16%。而混合特征算法的分类准确率分别为95.08%，86.87%；对溃疡型和蕈伞型食管癌进行分类，灰度共生矩阵的分类准确率分别为75.08%，66.96%；灰度直方图分类准确率分别为83.83%，77.23%。而混合特征算法的分类准确率分别为80.98%，73.66%。结论 灰度直方图特征的分类准确率比灰度共生矩阵特征的平均高10%，混合特征更适合于溃疡型，缩窄型食管癌的分类。而灰度直方图特征更适合于溃疡型，蕈伞型食管癌的分类；RF的分类能力比C4.5决策树高。此算法可为X线食管造影图像的分类提供参考。

Research on RF and C4.5 Decision Tree in Image Classification of Esophageal Cancer

Objective To explore the application of RF and C4.5 decision tree to the classification of X-ray esophageal images and to verify the classifier's ability to classify texture features.Methods From January to June 2018, the radiologists of the first affiliated Hospital, the second affiliated Hospital and the third affiliated (tumor) Hospital of Xinjiang Medical University selected 560 X-ray images of ulcerative, constrictive and mushroom esophageal cancer to extract the gray level symbiosis matrix. Grayscale histogram and mixed feature; RF and C4.5 decision tree are used to study the classification by adjusting the parameters.Results RF and C4.5 decision tree were used to classify ulcerative and constricted esophageal cancer. The classification accuracy of gray co-occurrence matrix was 73.30%and 67.76%.The classification accuracy of gray histogram was 84.55% and 76.16%,respectively.The classification accuracy of comprehensive feature algorithm was 95.08% and 86.87%, the classification accuracy of ulcerative and mushroom esophageal cancer was 75.08% and 66.96%, respectively, and the classification accuracy of gray histogram was 83.83%and 77.23%, respectively. The classification accuracy of comprehensive feature algorithm was 80.98% and 73.66%,respectively.Conclusion The classification accuracy of grayscale histogram is 10% higher than that of gray level co-occurrence matrix. The comprehensive feature is more suitable for classification of ulcerative and constrictive esophageal cancer. The gray histogram features are more suitable for the classification of ulcerative and mushroom esophageal cancer, and the classification ability of RF is higher than that of C4.5 decision tree. This algorithm can provide reference for the classification of X-ray esophageal images.