Automatic detection of cultured cells for robotic micromanipulation systems

Recent progress in the development of methods for molecular genetic analysis has brought sensitivities to the level where single cells can be analyzed. However, to carry out assays on significant numbers of cells, high throughput robotic systems for automatic cell micromanipulation are needed. This thesis project is part of the NIH-sponsored project entitled "Robotic Preparation of cDNA from Single Cells", whose objective is to develop a system that will facilitate analysis of gene expression in single, viable cells selected on the basis of multidimensional microscopy. In order to achieve this objective, a number of issues concerning the automatic detection of cultured cells in digital images obtained with transmitted light illumination are investigated in this thesis.;To extract features that are better suitable for classification tasks, a novel strategy for combining Fisher's Linear Discriminant (FLD) preprocessing with a feed-forward neural network for distinguishing "Cell" and "Non-cell" objects is first proposed. This technique is applied to various experimental scenarios utilizing different imaging conditions and the results are compared with those for the traditional Principal Component Analysis (PCA) preprocessing.;To move forward towards the goal of a practical automatic cell detection system, the problem of discriminating between "Viable cells" and "Non viable-cells" is formulated as a supervised, binary pattern recognition problem and solved using a Support Vector Machine (SVM) with an improved training algorithm proposed in this thesis (Compensatory Iterative Sample Selection, CISS). The new training algorithm is designed to solve the imbalanced large training set problem, which represents a difficult challenge for SVMs. It is also systematically studied under various class-size ratios and overlap conditions and found to outperform several commonly used methods, primarily owing to its ability to choose the most representative samples for the decision boundary.;The binary classification problems are further extended to facilitate detection of multiple cell types in mixtures. This task is formulated as a supervised, multiclass pattern recognition problem and solved by extension of the Error Correcting Output Coding (ECOC) method to enable probability estimation. The use of probability estimation provides both cell type identification as well as cell localization relative to pixel coordinates. This approach has been systematically studied under different overlap conditions and produced sufficient speed and accuracy for use in some practical systems.;In the present work, the multiclass cell detection framework is also extended to composite images consisting of images obtained with three different contrast methods in transmitted light. The use of multiple contrast methods improves the detection accuracy over single ones since it introduces more discriminatory information into the system. With regard to the composite images, Kernel PCA preprocessing is found to be superior to traditional linear PCA preprocessing, primarily owing to the fact that Kernel PCA can capture high-order, nonlinear correlations in the high dimensional image space.