基于形态特征判别超声图像中乳腺肿瘤的良恶性

乳腺肿瘤超声图像的形态特征对判别肿瘤的良恶性具有重要的价值。为提高乳腺肿瘤超声诊断的准确率,提出一种基于其形态特征进行分类判别的计算机辅助诊断系统。该系统首先采用灰度阈值分割和动态规划相结合的方法提取超声图像中乳腺肿瘤的边缘,然后对所得边缘计算相应的三种形态参数,最后分别采用Fisher线性判据、误差反向传播神经网络和径向基函数神经网络对形态参数进行分类。该系统在157幅乳腺肿瘤(包括良性81例、恶性76例)超声图像上训练和测试,三种分类器均能取得较高的判别精度,其中误差反向传播神经网络和径向基函数神经网络的判别准确率、敏感性和特异性分别高达94.95 %、95.74%和94.23%。结果表明,基于乳腺肿瘤超声图像的形态特征建立的神经网络系统对肿瘤的良恶性具有较好的判别能力。

Breast tumor classification based on shape features of ultrasonic images

The shape features of breast tumor in ultrasonic images are of great significance in the diagnosis of breast cancer.A computer-aided diagnosis system based on shape features was proposed to increase the accuracy of ultrasonic diagnosis of breast tumors.The tumor boundaries were firstly obtained using the gray-level threshold segmentation and dynamic programming and three shape features were subsequently calculated.Finally,the Fisher linear discriminant,neural network with error back propagation algorithm and radial basis function network were applied respectively to classify breast tumors as benign or malignant.Experiments on 157 cases(including 81 benign tumors and 76 malignant ones) show that all of three classifiers can achieve a higher precision,and the accuracy,sensitivity and specificity are as high as 94.95 %,95.74% and 94.23 % respectively for both of two neural networks.Therefore,it is concluded that the proposed system based on shape features performs well in the ultrasonic classification of breast tumors as benign or malignant.