Shape Particle Filtering for Image Segmentation

Abstract

Abstract. Deformable template models are valuable tools in medical image segmentation. Current methods elegantly incorporate global shape and appearance, but can not cope with localized appearance variations and rely on an assumption of Gaussian gray value distribution. Furthermore, initialization near the optimal solution is required. We propose a maximum likelihood shape inference that is based on pixel classification, so that local and non-linear intensity variations are dealt with naturally, while a global shape model ensures a consistent segmentation. Optimization by stochastic sampling removes the need for accurate initialization. The method is demonstrated on three different medical image segmentation problems: vertebra segmentation in spine radiographs, lung field segmentation in thorax X rays, and delineation of the myocardium of the left ventricle in MRI slices. Accurate results were obtained in all tasks.

Unifying Statistical Classification and Geodesic Active Regions for Segmentation of Cardiac MRI

“…They report a point to contour distance of 1.5 pixels and an intraobserver variability of 1 pixel, evaluated on eight slices from two test subjects in cine short-axis MRI data. de Bruijne and Nielsen [10] have presented an automatic segmentation method using shape particle filtering based on point distribution models (PDMs) that require point correspondences during the training phase. The method is evaluated on a data set consisting of 14 short-axis enddiastolic cardiac MRI slices with manually placed landmarks on the epi-and endocardial contour, with similar results (mean distance to contour 1.1 ± 0.3 pixels).…”

“…This can be seen as an attempt to avoid having to solve the general correspondence problem [11], and can be advantageous for objects without well-defined anatomical landmarks, as is the case for many objects including the myocardium. Shape is inferred by shape particle filtering [10], which, like the region term, is based on the class probability maps from classification.…”

“…This can be realized in a variety of ways: we use shape particle filtering as in [10] since the method has demonstrated accurate results for shape inference based on kNN classification. The particle filtering does not easily get trapped in local maxima nor does it need initialization close to optimal solution.…”

“…Each hypothesis is associated with an image labeling that is compared to the label probability maps from classification. Particles are weighted by these likelihood terms for each region where c is a constant controlling the randomness of the process and m is the number of pixels inside the template [10]. A new set of N hypotheses are generated from the current set by random sampling proportional to the weights W f and random perturbation of duplicate particles, so that successful particles will multiply while unlikely shapes will vanish.…”

“…Active appearance models [9] utilize a linear model of appearance by applying principal component analysis (PCA) on the intensities inside an object, something that can produce unreliable results if the intensities are not Gaussian distributed. For application to data with irregular appearance, regional statistics need to account for local and nonlinear intensity variations [10].…”