乳腺肿瘤超声图像识别模式分类方法的比较研究

Research on pattern classification methods for ultrasound breast tumor image

目的利用乳腺肿瘤超声图像良恶性的不同特征,借助于模式分类方法对乳腺肿瘤良恶性进行识别,作为医生的计算机辅助诊断。方法本文研究基于乳腺肿瘤超声图像的原始特征参数已提取情况下,采用顺序前进搜索方法获得最优特征矢量,然后利用支撑矢量机、贝叶斯分类器、BP网络和Fisher线性判别器四种模式识别方法分别对乳腺肿瘤良恶性进行识别。结果基于200例病例随机划分为训练集100例和测试集100例进行测试,支撑矢量机、贝叶斯分类器、BP网络和Fisher线性判别器的Accuracy分别为0.960,0.940,0.932±0.013,0.930。结论支撑矢量机的分类性能优于其它分类器,能有效地对超声图像乳腺肿瘤进行良恶性识别。

Objective To develop a computer-aided diagnosis with multiple features based on Pattern classification methods to differentiate benign from malignant breast tumor. Methods In this paper, optimal feature vector was firstly obtained from features extracted from Ultrasound Breast Tumor Image using Sequential Forward Selection Algorithm, then four pattern classification methods was used to classify the breast tumor, these pattern classification methods include SVM, BP, Bayes and Fisher classifier. Results Experiments on 200 ultrasonic images, randomly divided into training set 100 and prediction set 100, showed that the Accuracy of SVM, Bayes, BP and Fisher was 0.960, 0.940, 0.932±0.013, 0.930 respectively. Conclusion SVM classifier has the best performance and can effectively differentiate benign and malignant lesions.