Feature extraction from biological images with a focus on moment invariants

Bulent Bayraktar, Purdue University

Abstract

This research utilizes tools from pattern recognition for feature extraction in shape (e.g., morphology) and content (e.g., texture) analysis of biological images. The focus is on the theory of moments and the applications of both continuous and discrete moment invariants, which can quantify not just objects with well-defined boundaries but contents of more complex shaped objects without any clear boundaries. ^ In this work, moment invariants are used in the investigation of the effects of resolution changes (in biological imaging systems) on the descriptors of cell shape in terms of stability and consistence. The results show that the discrete-domain orthogonal Krawtchouk moment invariants are better cell-shape descriptors than are continuous-domain non-orthogonal geometric moment invariants in low-resolution images. ^ Using continuous-domain orthogonal Zernike moments, 6 different Listeria species are classified as pathogenic or non-pathogenic from their forward light-scattering images. These images are obtained by shining a 635-mm laser through pure bacterial colonies on agar plates in an optical setup. The colonies, along with the extracellular material they produce, show unique patterns that are captured using up to 20th order Zernike moment invariants. ^ A superior reconstruction scheme from discrete-domain orthogonal moments (Krawtchouk and Tchebichef moments) is proposed that provides a perfect (near-zero error) restoration of the original image from its moment decomposition. The algorithm uses an arbitrary-precision calculator for the calculation of polynomial basis coefficients for discrete moments. The arbitrary precision achieves the necessary precision and range for the storage and calculations of huge numbers that standard 64-bit IEEE floating-point arithmetic cannot provide in the case of larger images and high-order moments. ^ Changes in endothelial cell morphology and texture content are quantified under different shear stress conditions. In addition to traditional and simple descriptors, moment invariants and texture features are used for this study. The results show that we can reach conclusions about cell state without a full and precise segmentation. This novel method promises to form a new image analysis approach for high-throughput and high-content screening applications with vast amount of data.^

Degree

Ph.D.

Advisors

Peter C. Doerschuk, Purdue University, Joseph P. Robinson, Purdue University.

Subject Area

Engineering, Electronics and Electrical

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS