Johnston, Olivia; Preston, Kolten; Hoyt, Tyson; Owens May, April; Bentley, Kaden; Gunnerson, Shane; Johnson, Alex; Parr, McKenna; Reeves, Duncan; Parry, Whitney; Rawson, Clayton; Hart, Vern (Utah Valley University)
Faculty Advisor: Hart, Vern (Science, Physics)
Recent efforts in early cancer detection require identifying the disease at a cellular level, by distinguishing cancer cells from healthy cells at low concentrations (<0.1%). Cancerous cells typically have larger nuclei than healthy cells and can be distinguished using a variety of optical techniques, however, this process is complicated when the fraction of malignant cells is extremely low. As such, high-precision detection requires highly accurate measurements of cell confluence and the ratio of healthy to cancerous cells. Techniques such as machine learning and Fourier analysis have been used to auto-segment cells in microscopy images. However, these techniques often lack a ground truth standard to validate the segmentation results. We present a methodology for producing agarose tissue phantoms embedded with mixed polystyrene microbeads of varying diameters. These phantoms were imaged using a 2D translational stage and a microscope camera, collecting hundreds of images that were input to an artificially intelligent neural network for training and classification. The ability of this binary classifier to identify and quantify micro-beads in the images was assessed by comparing the automated results to manual counts, producing accuracies above 90% for bead sizes ranging from 50-200 microns. Auto-segmentation results will also be presented for mixtures of micro-beads and U-87 (glioblastoma) cancer cells, which differ in shape and morphology from the beads but whose boundaries are significantly less defined. The ability to accurately segment two different cell types in vitro would be highly beneficial for future cellular imaging studies.