V-Net Based Liver and Tumor Segmentation Using 3D ConditionalRandom Fields and Graph Cuts Method

Abstract

Automatic liver segmentation from 3D biomedical images is an essential task for many clinical applications and also helpful to radiologists and doctors to analyze and segment the liver tumor faster and accurately. However, it is still a challenging task due to the complex background of the liver, fuzzy boundary, and different appearance of liver and tumor. The present study was performed to propose more efficient deep learning based approach for liver and tumor segmentation from 3D CT scans using V-Net and two different post-processing methods: Conditional Random Field (CRF) and Graph Cuts. In propose approach, using 3-slice input and ResNet as base network outer performs by applying CRF as a post-processing step. Results are evaluated using different metrics that are used for biomedical images processing and with different combinations of parameters such as 3-slice input using ResNet with CRF and Graph Cuts. Results were significant by the proposed approach as compared to other state-of-the- art approaches. Tumor segmentation part are better than other approaches having 96.60%, 5.10%, -1.5%, 0.96mm, and 12.76mm for DSC, VOE, RVD, ASD, and MSD respectively for liver segmentation and 83.91%, 38.92%, -2.6%, 1.13mm and 5.76mm for DSC, VOE, RVD, ASD and MSD respectively for tumor segmentation.