Dense Depthwise Separable Convolution with Multi-Encoder Network for Liver Tumor Segmentation

Liver tumor segmentation from Computed Tomography (CT) images plays a significant role in diagnosis and treatment planning.However, automatic segmentation techniques face challenges because of irregularity, fuzzy boundaries and heterogeneity of tumor tissues.This research proposes Dense Depthwise Separable Convolution with Multi-Encoder Network (DDSC-MENet) for liver tumor segmentation.The MENet is developed in this research with down-sampling structure for liver tumor segmentation, which contains a whole down-sampling path with four encoders along with the same structures.The decoder connections in MENet allow the model to obtain feature representations from multiple encoder branches and perform integration of multi-scale and multi-modal information.The DDSC-MENet enhances the capacity of segmentation of tumors having different shape, sizes or intensity to the existing regular model.The Convolution Neural Network with Multiple Parametric Exponential Linear Units (CNN-MPELU) is applied in the classification stage for classifying liver tumors, which helps to resolve the vanishing gradient issues and quickens the convergence.The DDSC-MENet achieves dice of 94.58% and 92.62% for LiTS2017 and 3DIRCADB datasets respectively, which is better than Multi-Scale Feature Attention Network (MS-FANet).