基于磁共振图像的脑瘤MGMT表达状况检测算法

Detecting MGMT expression status of glioma with magnetic resonance image

针对脑胶质瘤的O6-甲基鸟嘌呤-DNA甲基转移酶(MGMT)表达状况检测受主观影响的问题,以中国脑胶质瘤患者的磁共振图像(MRI)为研究对象,提出一种包括特征提取、特征优化和模式分类的图像处理方法.利用图像的灰度共生矩阵、灰度-梯度共生矩阵和二维离散正交S变换(2D-DOST)提取肿瘤病变区域的纹理特征,结合环形增强和年龄特征构成初始特征集.将k最邻近法(KNN)与支持向量机(SVM)结合,进行特征优化.使用留一交叉检验法(LOOCV),将最优特征集进行SVM分类.分别对25位脑胶质瘤患者的T1加权、T1增强和FLAIR序列的磁共振图像进行分析.结果表明,该算法能够降低特征集的冗余程度,克服小样本分类困难,准确有效地检测MGMT表达状况.

In order to overcome the deficiency of strong subjectivity in detecting O6-methylguanine-DNA methyltransferase（MGMT） expression of gliomas,an image processing method was proposed to analyze the magnetic resonance images（MRI） of Chinese glioma patients.The method included feature extraction,feature optimization and pattern recognition.Gray co-occurrence matrix,gray level-gradient co-occurrence matrix and two-dimensional discrete orthogonal S-transform（2D-DOST） were utilized to extract the texture features in the tumor area.Ring enhancement and age were also added in the initial feature set.Then k-nearest neighbor（KNN） and support vector machine（SVM） were combined to search optimal features.The optimal feature set was classified by SVM in a leave-one-out cross validation strategy（LOOCV）.T1-weighted,T1-enhanced and FLAIR MRI of 25 glioma patients were analyzed.Results show that the algorithm can reduce the redundance of feature set,overcome the difficulty of small sample classification and identify the status of MGMT expression accurately and effectively.