Classification of Breast Tumors on Ultrasound Images Using a Hybrid Neural Network
Ling Zhong, Jiangli Lin, Deyu Li, Tianfu Wang, Yulan Peng, Yan Li Luo · 2007
Classification of breast tumors through the contour complexity parameter estimated by divider-step method was studied, using a hybrid neural network- the combination of an unsupervised self-organizing mapping network (SOM) and a multilayer perception (MLP) network with error back- propagation (BP) algorithm. The SOM was used to identify clusters and their centers in data (259 cases). Two-cluster data was then obtained by K-Nearest Neighbor. A profile for each cluster was determined by specified distance from its center. The cluster "profile" provided typical cases in the cluster and was applied to BP-ANN as the training set. The 96% specificity at 91.8% sensitivity was achieved after training. The results show the hybrid neural network is capable to produce good performance without labels by small training set.