Relaxed Conditional Hierarchical Statistical Shape Model of Multiple Organs

This paper proposes a relaxed conditional hierarchical statistical shape model (SSM) of multiple organs. After extracting shape and pose parameters from the training label dataset of multiple organs, the shape model and the pose model of each organ are constructed by principal component analysis (PCA). Subsequently, the principal scores of all organs are concatenated into a vector, and the vectors computed from the training dataset are forwarded to the PCA-based statistical modeling of the multiple organs under conditions of their neighboring organs. A relaxation scheme is introduced, to take into account errors in the conditions. This study focuses on modeling of a spleen and a gallbladder given a liver as a conditional organ. The performance of the model was evaluated with the measures of generalization and specificity, which were computed by three-fold cross-validation using labels of 27 abdominal CT volumes with the size of 170 × 170 × 110 voxels and a resolution of 1.8809 mm/voxel. Compared with a hierarchical SSM without conditions, generalization and specificity were improved from 0.488 to 0.506 and from 0.319 to 0.328 on average, respectively. In addition, the proposed relaxed conditional hierarchical SSM outperformed a hierarchical SSM with hard conditions. The performance indices were improved by 0.040 and 0.010 for generalization and specificity, respectively.