Experiments with nonlinear discriminants in the analysis of fine needle aspirates
Abstract
Fine Needle Aspiration (FNA) cytology provides a noninvasive and inexpensive procedure for diagnosing breast tumors. In contrast to conventional biopsies which involve surgery, in FNA a cytological sample of the tumor is aspirated using a needle. Since the architecture of the tumor cannot be inferred from a cytological sample, diagnosing a tumor using FNA has turned out to be a challenging task; the accuracy of visual diagnosis (by pathologists) of malignant tumors is reported to vary from 65% to 98%. To remedy the large variation in diagnostic accuracy, in recent years several research groups have attempted to automate the diagnostic procedure using pattern classification algorithms. The research reported in this dissertation contributes to the ongoing effort to automate FNA diagnosis. Previous research has focused largely on piecewise linear discriminants for classifying tumor samples. While piecewise linear discriminants are computationally tractable, it is not clear that they provide an acceptably accurate diagnostic procedure. The premise of this dissertation is that both the diagnostic accuracy and the variance in diagnosis, for malignant tumors, improve with increasing nonlinearity of the discriminants. Hence we studied discriminants with progressively increasing nonlinearity. The discriminant schemes we devised involved constrained mixed integer nonlinear optimization problems that are computationally formidable. We employed unconventional transformations, that could be of independent interest in nonlinear optimization, to transform the constrained mixed integer nonlinear programs into unconstrained nonlinear programs that are computationally tractable. We then studied the performance of the various discriminants on the Wisconsin Breast Cancer Database. Results of our computational experiments suggest that the reliability of diagnosis of malignant tumors does improve with increasing nonlinearity of the discriminants.
Degree
Ph.D.
Advisors
Prabhu, Purdue University.
Subject Area
Biomedical research|Pathology
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.