Image classification of unlabeled malaria parasites in red blood cells

This thesis presents a method to automatically detect unlabelled malaria parasites in red blood cells. The current approach widely used to diagnose malaria is via microscopic examination of thick blood smear which is a time consuming process requiring extensive training. The goal is to develop an automate process to identify malaria infected red blood cells to reduce the time and training expenses incurred to diagnose malaria. Major issues in automated analysis of microscopy images of unstained blood smears include overlapping cells and oddly shaped cells from the sample preparation process. Robust templates were created to label cells as infected or uninfected. From these templates, classifiers were trained offline. Different feature descriptors such as PCA, HOG and GIST were compared. HOG performed better than the other two methods for this application. The infected cells were detected using a two stage approach. Viola-Jones object detection framework was applied to extract image regions with to detect red blood cells from background in the first stage. In the second stage, infected cells were detected by combining another Viola-Jones object detector with their morphological features extracted using intensity thresholding, colour thresholding and circular Hough transform. Results show our approach out-performs classification approaches with PCA features by 50% and cell detection algorithms with Hough transforms by 24%. Graphical user interfaces are built for convenient employment of this approach and to improve template database create. Known related work is applied on images where the malaria parasites were stained, where as in this work, the parasites are not stained. It is more challenging to design algorithms for unstained parasites as the differences between infected cells and uninfected cells are less obvious. However, this approach will allow observations of malaria parasite progression and provide statistics on how malaria parasites respond to different drug treatments over time.Master of Science (Computer Control and Automation