Construction of a probabilistic atlas for automated liver segmentation in non-contrast torso CT images

Abstract In this paper, we proposed a method to generate a probabilistic atlas of liver and use it for liver region segmentation from non-contrast torso CT images. The atlas was defined by two kinds of probability: liver location and its density (CT number). We developed a method to normalize the liver location in different patient cases by deforming the pre-segmented diaphragm and bone to a standard position and to generate a normalized likelihood image to present probability of liver location in torso CT image by projecting a large number of liver regions into a 3-D space. A Gaussian window function with dynamic parameter estimations was used to calculate the probability of liver density. 80 patient cases of non-contrast CT images with the pre-segmented liver regions were used for performance evaluation. We repeated the atlas generation and atlas-based segmentation of liver region based on non-contrast torso CT images using a leave-one-out method. The mean value of the coincidence ratio between the pre-segmented liver and atlas-based segmentation result was found to be 84%. The validation and usefulness of our atlas construction method were proved.