基于CT图像纹理分析的结肠息肉鉴别研究  

A study on the identification of colonic polyp based on texture analysis of CT image

作　　者：孟江 卢虹冰 徐肖攀 徐桓 张国鹏 梁正荣[3] MENG Jiang;LU Hong-bing;XU Xiao-pan(College of Biomedical Engineering, The Military Medical University of Air Force, Xi'an 710032, China)

机构地区：[1]空军军医大学生物医学工程学院,陕西西安710032 [2]中央军委后勤保障部卫生局药品仪器检验所,北京100071 [3]美国纽约州立大学石溪分校放射系,美国纽约11794

出　　处：《中国医学装备》2018年第3期10-14,共5页China Medical Equipment

基　　金：国家自然科学基金(81230035)"基于影像定量分析和特征可视化的虚拟内窥镜关键技术研究"

摘　　要：目的:从结肠CT影像及其高阶偏导图像中提取三维Haralick特征并进行特征寻优,以用于结肠息肉与正常肠壁组织的准确鉴别。方法:从111例确诊的结肠息肉患者的CT结肠镜(CTC)影像数据中共提取387个结肠息肉和387个正常的结肠壁组织作为三维感兴趣区域(ROI),并对其进行三维梯度和曲度变换。从每个ROI的CT图像、梯度图像和曲度图像中分别计算共生矩阵,并从这三类共生矩阵中分别提取60个Haralick纹理特征,共计180个特征。利用基于随机森林(RF)的特征选择策略进行特征寻优,并使用逻辑回归(LR)、支持向量机(SVM)、RF和K-近邻(KNN)4种不同的分类器验证最优子集的分类性能。结果:对180个特征进行特征筛选,获得37个特征作为最优特征子集,利用最优特征子集与4种不同分类器进行验证,其区分肠壁和息肉的平均灵敏度均在0.99%以上,平均特异度均在0.98%以上,平均受试者工作特性曲线下面积(AUC)均为0.99。结论:基于灰度、灰度梯度和灰度曲度的纹理特征具有良好的敏感性和特异性,能够有效地鉴别结肠息肉与正常肠壁组织,对基于影像的结肠息肉临床筛查与诊断具有重要的参考价值与借鉴意义。Objective:To extract three dimensional(3D)Haralick characteristic from CT image of colon and its highorder partial derivative image and find out optimal characteristic so as to accurately differentiate polyp from normal wall tissue of colon.Methods:Totally387colonic polyps and387normal wall tissues of colon were extracted from the image data of CT colonoscope(CTC)of111patients with colonic polyp as3D volumes of interest(VOIs),and3D gradient and curvature maps were calculated and transformed on these VOIs.And co-occurrence matrices were calculated from CT image,gradient image and curvature image of each VOI.And60Haralick textural features were extracted from each kind of co-occurrence matrices,and totally180features were extracted from each VOI.The feature selection strategy based on random forest was adopted to find out optimal feature.Finally,the optimal subset was input into four different classifiers,which included of logistic regression(LR),support vector machine(SVM),random forest(RF)and K-near neighbour(KNN),respectively,to validate its classification performance in differentiating polyps from wall tissues.Results:37features that selected from180features were confirmed as the optimal feature subset.And the optimal feature subset were validated in four different classifier and the results indicated that the average sensitivity of identifying polyp from wall tissue was above99%,and the average specificity of that was above97%,and all of the average area under the curve(AUC)of performance were0.99.Conclusion:The textural feature that based on gray level,gray gradient and gray curvature has better specificity and sensitivity,and it can effectively identify polyp and normal wall tissue in the colon.Therefore,it has importantly referential value for clinical screening and diagnosis of colonic polyp based on imaging.

关 键 词：结肠息肉 CT CT结肠镜 纹理分析 特征选择 三维感兴趣区域 

分 类 号：R735.3[医药卫生—肿瘤]