基于CT图像统计纹理特征的甲状腺结节识别技术

Statistic Texture Feature Based Thyroid Nodule Recognition on CT Images

目的探讨灰度共生矩阵和灰度梯度共生矩阵统计学纹理特征在CT图像上甲状腺结节良恶性鉴别的可行性。方法回顾性收集甲状腺结节经手术病理证实的CT图像134例,手动提取含结节的单侧甲状腺感兴趣区(region of interest,ROI)。计算ROI的统计学纹理特征并归一化到[0,1],支持向量机作为分类器,并结合留一交叉验证法来评价实验效果。结果统计学纹理特征在甲状腺结节良恶性鉴别中的准确率为0.76,敏感度0.60,特异性0.86和受试者操作曲线下面积为0.81。结论基于灰度共生矩阵和灰度梯度共生矩阵的统计法纹理特征,在甲状腺CT图像上对于结节的良恶性鉴别具有较好的分类效果。

Objective To evaluate the feasibility of classifying the malignant thyroid nodule from benign one on computed tomography（ CT） images based on gray level co-occurrence matrix and gray level gradient co-occurrence matrix texture features. Methods One hundred and thirty four CT images from inpatients underwent thyroid nodule surgery were enrolled in this study. A senior radiologist delineated the thyroid contour manually and segmented the region of interest（ ROI）. Texture features of GLCM and GLGCM were extracted and scaled to [0,1]. Support vector machine was used as classifier. Leave-one-out cross validation（ LOOCV） strategy was applied to evaluate the performance. Results Statistic texture features were applied on the classification of thyroid nodule recognition and the results of the proposed method were accuracy 0. 76,sensitivity 0. 60,specificity 0. 86 and area under receiver operating curve（ AUC） 0. 81 respectively. Conclusion The texture features of GLCM and GLGCM can be used as image biomarker in classifying malignant thyroid nodule from benign one on CT images.