Image cytometry classifiers : isolation of cancer cells from blood and morphometric characterization of cytotoxicity in high content screening

Advances in microscopy have made the automatic acquisition of large volumes of images from biological specimens a routine. In high content screening (HCS), databases of cytometric measurements from cell images can be easily made. The transformation of such cytometric data into biological knowledge is the general aim of this dissertation. One such transformation can help detect circulating tumor cells (CTCs) in blood based only on their morphology. CTCs travel from a primary tumor to settle in a metastatic site and form more tumors. Enabling the detection of CTCs can help characterize metastatic behavior and improve cancer diagnosis. Here, the use of neural networks in detecting CTCs based on nuclear morphology alone was explored (Chapter 1). Furthermore, various methods are proposed and tested to systematically improve the performance of such classifiers (Chapter 2). Additionally, since the method is novel, it could detect cells that may be key to the study of cancer progression but have been missed by current techniques. Therefore, it is also important to physically isolate the resultant cells to do down-stream experimentation. To accomplish this, a microscopy cell sorting instrument is developed and tested that retrieves desired cells directly from the surface of a microscopy slide with relatively high throughput (Chapter 3). Another transformation of cytometric data to biological knowledge was explored by attempting to gauge cytotoxicity using morphology alone. In HCS, hundreds of thousands of compounds are tested on cells and a desired biological response is assayed to find drug candidates. Assessing toxicity in that stage would stop the advancement of toxic candidates into more expensive testing and clinical trials. Implementing cytotoxicity assays into screens, however, is prohibitively costly. A method for mimicking the costly cytotoxicity measurements in large screens by using morphology alone is proposed and studied (Chapter 4). The method is then applied to a small screen and its performance is measured. Further, distortions often exist in screen data because of problems with preparation of microtiter plates. A median-based method has been developed and tested here to remedy this problem and make screen data statistically more viable (Chapter 5