Profile scale spaces for statistical image match in Bayesian segmentation.

摘要

Object boundaries in images often exhibit a complex greylevel appearance, and modeling of that greylevel appearance is important in Bayesian segmentation. Traditional image match models such as gradient magnitude or static templates are insufficient to model complex and variable appearance at the object boundary; in the presence of image noise; jitter in correspondence, and variability in a population of objects.; I present a new image match model for Bayesian segmentation that is statistical, multiscale, and uses a non-Euclidean object-intrinsic coordinate system. The segmentation framework is based on the spherical harmonics object representation and segmentation framework of Kelemen et al., which in turn uses the profile-based image match model of Active Shape Models. The traditional profile model does not take advantage of the expected high degree of correlation between adjacent profiles along the boundary. My new multiscale image match model uses a profile scale space, which blurs along the boundary but not across the boundary. This blurring is done not in Euclidean space but in an object-intrinsic coordinate system provided by the geometric representation of the object. Blurring is done on the sphere via a spherical harmonic decomposition; thus, spherical harmonics are used both in the geometric representation as well as the image profile representation. The profile scale space is sampled after the fashion of the Laplacian pyramid; the resulting tree of features is used to build a Markov Random Field probability distribution for Bayesian image match.; Results are shown on a large dataset of 114 segmented caudates in T1-weighted magnetic resonance images (MRI). The image match model is evaluated on the basis of generalizability, specificity, and variance: it is compared against the traditional single scale profile model. The image match model is also evaluated in the context of a full segmentation framework, when optimized together with a shape prior. I test whether automatic segmentations using my multiscale profile model come closer to the manual expert segmentations than automatic segmentations using the single-scale profile model do. Results are compared against intra-rater and inter-rater reliability of manual segmentations.