A Statistical Appearance Model Based on Intensity Quantile Histograms

Abstract

We present a novel histogram method for statistically characterizing the appearance of deformable models. In deformable model segmentation, appearance models measure the likelihood of an object given a target image. To determine this likelihood we compute pixel intensity quantile histograms of object-relative image regions from a weighted 3D image volume near the object boundary. We use a Gaussian model to statistically characterize the variation of histograms understood in Euclidean space via the Mallows distance. The probability of gas and bone tissue intensities are separately modeled to leverage a priori information on their expected distributions. The method is illustrated and evaluated in a segmentation study on CT images of the human left kidney. Results show improvement over a profile based appearance model and that the global maximum of the MAP estimate gives clinically acceptable segmentations in almost all of the cases studied.

Clustering on Local Appearance for Deformable Model Segmentation

“…We use volume overlap (intersection over average) with the expert manual segmentations to compare our boundaries to boundaries computed using the global exterior image match described in [6]. Segmentations using our local regions appearance scheme improve results in 57.5% (46 of 80) of bladders and 53.8% (43 of 80) of prostates over all patients.…”

“…In [6] we describe an approach to image match to probabilistically represent the appearance of an object in an image. Appearance is in the form of regional intensity quantile functions, derived from intensity histograms within objectrelative regions, such as the interior near the object boundary.…”

“…Some of the resulting intensity models are trained on summary statistics such as mean and variance [4], which themselves capture limited information. Other recently developed models use the entire intensity distribution, and either compute an image match with respect to a single reference distribution [5], or most recently, train the variation in the distribution and compute a real probability to use as the likelihood p(I|m) [6].…”

“…In [6] we use two object-relative regions: interior and exterior to the object boundary. We sample from the regions and convert the resulting distributions to regional intensity quantile functions (RIQFs), which are amenable to linear statistical methods such as principal component analysis (PCA).…”

“…In this paper, we extend the work of [6] to account for the diverse local intensity distributions found on the exterior of an object of interest. We capture intensity quantile functions at many local regions outside the object boundary in training and cluster on the pooled data, deciding the number of sources, or region types.…”