Recognition of Breast Ultrasound Images Using A Hybrid Method
Kai Zheng, Tianfu Wang, Jiangli Lin, Deyu Li · 2007
Breast cancer is one of the most common cancer in women. A novel method is presented in this paper for classification of the breast tumors as benign or malignant. The method combines k-means classification and a multilayer perception network with error back-propagation (BP) algorithm. The k-means which is an unsupervised classification method is used to get the cluster centers and select the training samples. Only the samples within a specified distance from the cluster centers could be selected as our training samples. The BP neural network is used to train the samples and recognize the tumors. The fractal dimension of an ultrasound image of breast tumor is extracted as our feature, which is estimated with the difference between gray values of neighboring pixels by using the fractal Brownian motion. Experiments are done on 125 benign tumors and 110 malignant ones. The recognition rate of the malignant tumors is 94.5% while that of the benign ones is 93.6%. The result shows that the proposed method can classify the breast tumors effectively.