Abstract
Training a shape prior has been potent scheme for anatomical object segmentations, especially for images with noisy or weak intensity patterns. When the shape representation lives in a high dimensional space, principal component analysis is often used to calculate a low dimensional variation subspace from frequently limited number of training samples. However, the eigenmodes of the sub-space tend to keep the large-scale variation of the shape only, losing the detailed localized variability which is crucial to accurate segmentations. In this paper, we propose a large-to-fine-scale shape prior for probabilistic segmentation to enable local refinement, using a deformable medial representation, called the m-rep. Tests on the goodness of the shape prior are carried out on large simulated data sets of a) 1000 deformed ellipsoids with mixed global deformations and local perturbation; b) 500 simulated hippocampus models. The predictability of the shape priors are evaluated and compared by a squared correlations metric and the volume overlap measurement against different training sample sizes. The improved robustness achieved by the large-to-fine-scale strategy is demonstrated, especially for low sample size applications. Finally, posterior 3D segmentations of the bladder from CT images from multiple patients in day-to-day adaptive radiation therapy demonstrate that the local residual statistics introduced by this method improve the segmentation accuracy.