Abstract
A blood cell count is a common diagnostic tool in medicine, and one way to obtain such a count is from an image of a blood smear. Researchers at the Center for Biophotonics Science and Technology (CBST) at the University of California, Davis have developed an attachment to convert a cell phone to a microscope. The images provided by this cell-phone microscope suffer from several artifacts, such as radial distortion and non-uniform illumination. It is desired to develop a software application for a smart phone to perform image processing and pattern recognition that can return an approximate blood count. In this work, prototype software has been developed on a personal computer (PC) that performs the whole procedure of image processing and pattern recognition to provide an approximate red blood cell count. To do the red blood cell count, images that are taken of a blood sample by a smart phone are transferred to a PC for processing. Radial distortion correction and cropping the defocused area of the image are done as pre-processing steps in preparation for robust cell recognition. Adaptive multi-level segmentation is performed as the second step to transform the image to a fuzzy scene, followed by the red cell recognition step. A fuzzy approach is taken for red cell recognition. The fuzzy approach presented in this work utilized fuzzy sets and not fuzzy logic. Adaptive image fuzzification and fuzzy criterion functions proposed in this thesis have higher performance than conventional counting methods. The proposed approach is robust against fuzziness of the image due to the poor quality of a cell phone image, taken under non-laboratory conditions. The recognition process in this application is a blind search method that is independent of manual calibration and learning. Most of this work has been dedicated to enhancing the algorithm of cell recognition even in poorly-illuminated images. This work focuses on red blood cell counting. However, the concept can be extended to other blood smear counting, such as white blood cells and platelets. This algorithm is tested on seven blood smear images, and the average values for precision and recall are 95.6 percent and 95.4 percent, respectively.