Influence of Voxel-Connection Structure in Organ Segmentation Based on Conditional Random Field

摘要

In this paper, we investigate the influence of voxel-connection structure in torso organ segmentation from CT volumes based on conditional random field (CRF). Many methods of automated segementation from CT volumes have been proposed. However, lots of parameters require to be adjusted empirically to obtain precise organ regions in these methods. Here, we extract organ regions from CT volumes by estimating the labels of each voxel based on CRF model. We construct CRF model using the stochastic gradient descent algorithm in the learning phase and maximum a posteriori (MAP) inference in the prediction of the model. To evaluate the impact of voxel-connection structure on CRF-based organ segmentation, we perform the experiments by CRF model with new voxel-connection structure. Furthermore, we investigate the influence caused by the number of training loops and the number of iterations of parameter updating as well. The experimental results of 10 CT volumes showed that the connection structure, the number of loops and iteration have influence on the performance of organ segmentation based on CRF model.