TABLE OF CONTENTS MINGZHOU SONG
Department of Computer Science, New Mexico State University
P.O. Box 30001, MSG GS, Las Cruces, NM 88003, USA
joemsong@cs.nmsu.edu
ROBERT M. HARALICK
Doctoral Program in Computer Science, Graduate Center, City University of New York
365 Fifth Ave., New York, NY 10016, USA
haralick@gc.cuny.edu
We describe a non-parametric pixel appearance probability model to represent local image information. It allows an optimal image analysis framework that integrates low-and high-level stages to substantially improve overall accuracy of object reconstruction. In this framework, feature detection would be an overall consequence rather than an intermediate result. The pixel appearance probability model is a probability density function obtained by grid quantization. A grid is found by a genetic algorithm and a local refinement algorithm. The density values in each cell of the grid are computed by smoothing neighboring cells. We apply the pixel appearance probability model to represent features of echocardiographic images. We illustrate the substantially improved performance on left ventricle surface reconstruction due to the proposed model.
Keywords: Non-parametric pixel appearance probability model; grid quantization; ultrasound imaging.
The ultimate goal of medical image analysis is to acquire quantitative representations of objects that are of medical concern from observed images. In echocardiography, one objective is to create a three-dimensional (3D) Left Ventricle (LV) surface model, including the EPIcardium (EPI), the outer surface of the LV, and the ENDOcardium (ENDO), the inner surface of the LV. A standard two-stage approach comprises feature detection and object reconstruction. Feature detection classifies each pixel into...
