Abstract
This paper proposes a semi-automated organ segmentation algorithm using landmarks (LMs) from a computed tomography (CT) volume. Here, several LMs are manually defined by the user on a target organ's surface in an input volume. A patient-specific probabilistic atlas (PA) is constructed based on sparse representation obtained using training labels computed by minimizing the square error between the input and training label LMs as well as the L1 norm term. This paper presents the experimental results for the purpose of segmentation of the liver, gallbladder, and right and left kidneys, in which the proposed PA was proved to be effective in an organ of irregular shape. The average Jaccard Indices of the maximum a posteriori followed by graph cuts-based segmentation of all the target organs were 0.757 for a conventional PA, and 0.838 for the proposed PA, respectively. It was statistically confirmed that the segmentation performance using a proposed PA was superior to that using a conventional PA. We also discuss the pros and cons of the proposed PA by exploring the relationship between the organ's shape and segmentation performance.
