Triplanar Convolutional Neural Network for Automatic Liver and Tumor Image Segmentation

摘要

The automatic image segmentation of liver and liver tumors is important in the diagnosis and treatment of hepatocellular carcinoma. A novel triplanar fully convolutional neural network (FCN) composed of three 2D convolutional neural networks (CNNs) is proposed to handle the issue. It performs segmentation through the transverse plane, coronal plane, and sagittal plane and can effectively use multidimensional features for 3D segmentation. Then, a cascaded structure is used to balance the positive and negative samples for segmentation of the tumor. The experimental results are obtained through data analysis and tested on the 3DIRCADb. They show that our method outperforms the existing methods and achieves a volume overlap error of 6.7% and 3.6% on the liver and tumors respectively.