Fast graph-based medical image segmentation with expert guided statistical information

In radiotherapy treatment planning, delineation of normal organs at risk in images is one of the most time-consuming tasks carried out routinely by human experts. Previously we proposed a speedy semi-automatic segmentation method based on a statistical graphical model, Conditional Random Field (CRF,) from which an energy function is defined to obtain Maximum-a-posteriori (MAP) estimation of the segmentation via a fast graph cut algorithm. The probabilistic regional and boundary terms in the energy function are estimated from the training samples collected locally from the the human expert via interactive tool or a training database. In this paper, we present a simple acceleration technique that dramatically improves the speed without sacrificing the accuracy of the segmentation. In the context of slice-by-slice medical image segmentation, we accelerate the process by partially reusing the graph constructed from a previous segmented slice based on the likeness of two consecutive images. Experiment results in 5 liver cases show differences between the manually segmented volumes and our estimated volumes were less than 5%. The differences are within the normal variation of manual segmentation from inter- and intra-observers. Accelerated segmentations show no degradation in terms of accuracy compared to full segmentations. The computation time per slice is within 300 millisecond CPU time for full segmentation and 110 millisecond for accelerated segmentation. The semi-automatic segmentation method proposed achieves similar segmentation done by human expert in significantly lesser time while preserving the human oversight required during the treatment planning process.