Machine Learning Study of Several Classifiers Trained With Texture Analysis Features to Differentiate Benign from Malignant Soft-Tissue Tumors in T1-MRI Images
Machine Learning Study of Several Classifiers Trained With Texture Analysis Features to Differentiate Benign from Malignant Soft-Tissue Tumors in T1-MRI Images
复制标题
DOI：
10.1002/jmri.22095
复制
发表时间：
2010-03-01
影响因子：
4.4
通讯作者：
Van Dyck, Dirk
中科院分区：
文献类型：
--
作者：
Juntu, Jaber;Sijbers, Jan;Van Dyck, Dirk
目的：为了研究，从机器学习的角度来看，几个机器学习分类器的性能，使用纹理分析的非增强T1-MRI图像中的软组织肿瘤提取的功能，以区分恶性和良性tumors.Materials和方法：纹理分析功能提取的肿瘤区域从T1-MRI图像的49例恶性和86例良性软组织肿瘤的临床证明。三个传统的机器学习分类器进行了训练和测试。最好的分类器进行了比较，放射科医生通过McNemar的统计test.Results：SVM分类器的性能优于神经网络和C4.5决策树的基础上，分析其受试者工作曲线（ROC）和成本曲线。SVM的分类准确率为93%（91%特异性; 94%灵敏度），优于放射科医师的分类准确率90%（92%特异性：81%灵敏度）。结论：使用纹理分析特征训练的机器学习分类器对于检测T1-MRI图像中的恶性肿瘤具有潜在价值。对分类器的学习曲线的分析表明，小于100个T1-MRI图像的训练数据大小足以训练机器学习分类器，其表现与放射科专家一样好。
Purpose: To study, from a machine learning perspective, the performance of several machine learning classifiers that use texture analysis features extracted from soft-tissue tumors in nonenhanced T1-MRI images to discriminate between malignant and benign tumors.Materials and Methods: Texture analysis features were extracted from the tumor regions from T1-MRI images of clinically proven cases of 49 malignant and 86 benign soft-tissue tumors. Three conventional machine learning classifiers were trained and tested. The best classifier was compared to the radiologists by means of the McNemar's statistical test.Results: The SVM classifier performs better than the neural network and the C4.5 decision tree based on the analysis of their receiver operating curves (ROC) and cost curves. The classification accuracy of the SVM, which was 93% (91% specificity; 94% sensitivity), was better than the radiologist classification accuracy of 90% (92% specificity: 81% sensitivity).Conclusion: Machine learning classifiers trained with texture analysis features are potentially valuable for detecting malignant tumors in T1-MRI images. Analysis of the learning curves of the classifiers showed that a training data size smaller than 100 T1-MRI images is sufficient to train a machine learning classifier that performs as well as expert radiologists.