关键词: 灰度共生矩阵, 甲状腺核磁共振图像, Logistic回归模型, ROC曲线下面积

Abstract: To improve the predictive accuracy of the lesions and the normal areas of the thyroid MR images, an adaptive window sizes approach is proposed based on Gray Level Co-occurrence Matrix（GLCM） texture feature. Firstly, it considers the effect of the gradient information of ROIs on sliding window size set when it computes GLCM and extracts the gradient information of ROIs. Then, it sets different sliding window size on the gradient. Moreover, it obtains the mean of the co-occurrence matrix in four directions in order to eliminate the anisotropy. Finally, it computes the mean and variance of correlation, energy, contrast, homogeneity and entropy in GLCM. In this study, 94 cases of thyroid images are used to predict the classification accuracy by using logistic regression model analysis. The experimental result shows that the improved method outperforms other methods, and has the better diagnostic performance with an accuracy of 96.8%, a sensitivity of 97.9%, a specificity of 95.7% and an AUC of 0.968. The experimental result indicates that the feature extraction method can effectively assist the doctors to make the correct diagnosis of thyroid MR images.

Key words: Gray Level Co-occurrence Matrix（GLCM）, thyroid magnetic resonance, logistic regression model, area under ROC curve