One-Class Artificial Neural Network Based Liver Vessel Segmentation

Blood vessel segmentation is a greatly significant task in medical image processing domain. Segmenting liver vascular is especially full of challenges because of its complex structure. In this paper, we propose a one-class artificial neural network combining with 3D Perona-Malik (PM) equation for liver segmentation. The essential processes of designed method mainly contain data preprocessing by window level and center adjustment, feature extraction by PM equation and pixel-wise classification by proposed neural network. The experiment shows window level and center adjustment and PM equation are complementary and interacting. In addition, by applying Gaussian Mixture Model, it can be concluded that the window center value should be range from 150H to 200H which focuses on vessel part in the liver. In the paper, the iteration of PM equation is 10 times with parameters K, gamma, step equaling to 20, 0.25, each volume’s spacing respectively. The class imbalance problem, which have a bad influence on classification-aimed artificial neural network, will be solved in the data preparation process. The designed one-class ANN model works fast and get a decent result. The testing accuracy of a CT volume is 93%.