Supervoxels based Graph Cut for Medical Organ Segmentation

Abstract

Organ segmentation is one of the most fundamental and challenging tasks in computer aid diagnosis system. Researches successfully working on interactive segmentation for medical image include graph cuts and random walk. However, graph cut-based organ segmentation for 3D medical volume data requires an optimization procedure of cutting the object/background regions on a very large-scale graph, which not only consumes large amount of memory and but also requires expensive computational cost. This paper conquers the drawbacks of graph cut-based organ segmentation via instead of voxel with supervoxel as nodes for constructing graph in interactive 3D organ segmentation that can greatly reduces node number and connected edges; voxels of medical data with similar intensity magnitude and near spatial relation are grouped into the supervoxels and such supervoxels are used as pseudo nodes of graph for cutting object (organ) and background regions, named as supervoxel-based graph cut. To validate the effectiveness and efficiency of the proposed method, we conduct experiments on 10 medical data, which possibly have tumors inside organ, or have abnormal deformed organ shape. The experimental results show that the proposed method is much superior than conventional graph cut-based method in term of accuracy, computational time and memory usage.