A knowledge-based interactive liver segmentation using random walks

Abstract

A random walks-based (RW) segmentation method has been gaining popularity in recent years with its ability to interactively segment the objects with minimal guidance. It has potential applications in segmenting the 3D image. However, due to the large computational burden of the classical RW algorithm, it is a challenge to use this algorithm to segment 3D medical images interactively. Hence, a knowledge-based segmentation framework for the liver is proposed based on random walks and narrow band threshold (RWNBT). Our strategy is to employ the previous segmented slice to achieve a prior knowledge (the shape and intensity constraints) of liver for automatic segmentation of the adjacent slice. With a small number of user-defined seeds, we can obtain the segmentation results of the start slice in the volume which would be used as the prior knowledge of the segmented organ. According to this intensity constraints, the Candidate Pixels image can be generated by thresholding the organ models with Gaussian Mixture Model (GMM), which can remove the noise and non-liver parts. Furthermore, the object/background seeds can be dynamically updated for the adjacent slice by combining a narrow band threshold (NBT) method and the shape constrains. Finally, a combinational random walker algorithm is applied to automatically segment the whole volume in a slice-by-slice manner. Comparing our method with conventional RW and the state-of-the-art interactive segmentation methods, our results show an improvement in the accuracy for liver segmentation.