A Simultaneous Segmentation and Reconstruction Model for ECT Images

Authors: 
Valen E. Johnson, James E. Bowsher
Duke University

Nov 30 1990

We propose a Bayesian model for the reconstruction of ECT images. The prior density effectively reduces the dimension of the estimation problem by merging pixel in the image with similar intensities, thus making EM type algorithms feasible. Two specific models are proposed; one imposes no prior constraint on the shapes of regions except that regions must be at least four contiguous pixels in size, and the other imposes minor shape constraints and requires that regions e at least five pixels in size. The model provides an automatic segmentation of the image which may be useful in quantization and visualization applications. Additionally, mean intensities and region sizes are immediately available as output from the algorithm. Computationally, the models are easy to implement and require approximately a 5% increase in computation over standard EM/MLE algorithms.

Keywords: 

Bayesian inference, Gibbs distributions, positron emission tomography, emission computer tomography, estimation-maximization algorithm, Bayesian data augmentation

Manuscript: 

PDF icon 1991-02.pdf