# Postmortem Liver Segmentation Algorithm with an SSM

To develop a segmentation algorithm with an SSM [182] for a postmortem liver, a similar framework for an in vivo liver [183] was employed, but differed in an SSM- and a MAP-based rough segmentation with a probabilistic atlas. A comparative study of algorithms using the aforementioned eight different SSMs is also given in this section.

# Method

The proposed liver segmentation algorithm consists of three steps: (1) rough segmentation, (2) SSM-based shape estimation, and (3) graph cuts with the estimated shape. The rough segmentation is performed by a probabilistic atlas-guided expectation maximization (EM) algorithm followed by a MAP segmentation [184], in which probability distribution of organsâ€™ features is assumed to be a mixture of Gaussians. The difference from the in vivo liver segmentation algorithm [183] is that the atlas-guided EM and MAP are repeated by updating the location of the probabilistic atlas according to the MAP segmentation result of the previous iteration. Such iteration is important to deal with the postmortem-specific deformation of organs. Subsequently, in the shape estimation process, the most similar shape to a MAP segmentation result is selected from an eigenshape space of an SSM. Finally, the graph cut-based segmentation with the estimated shape is performed. The energy function to be minimized is composed of a unary and pairwise terms. The unary term is a negative logarithm of posterior probability of liver given a CT volume in which a patient-specific probabilistic atlas calculated from the estimated shape is used as a prior probability. Pairwise terms consist of a conventional boundary term and a shape term that evaluates the difference in gradients between a segmented shape and the estimated one.