Abdallahi, H; (2016) Automatic analysis of malaria infected red blood cell digitized microscope images. Doctoral thesis , UCL (University College London).

Preview

Text Abdallahi_PhD 2016 ABDALLAHI, H.pdf Download (6MB) | Preview

Abstract

Malaria is one of the three most serious diseases worldwide, affecting millions each year, mainly in the tropics where the most serious illnesses are caused by Plasmodium falciparum. This thesis is concerned with the automatic analysis of images of microscope slides of Giemsa stained thin-films of such malaria infected blood so as to segment red-blood cells (RBCs) from the background plasma, to accurately and reliably count the cells, identify those that were infected with a parasite, and thus to determine the degree of infection or parasitemia. Unsupervised techniques were used throughout owing to the difficulty of obtaining large quantities of training data annotated by experts, in particular for total RBC counts. The first two aims were met by optimisation of Fisher discriminants. For RBC segmentation, a well-known iterative thresholding method due originally to Otsu (1979) was used for scalar features such as the image intensity and a novel extension of the algorithm developed for multi-dimensional, colour data. Performance of the algorithms was evaluated and compared via ROC analysis and their convergence properties studied. Ways of characterising the variability of the image data and, if necessary of mitigating it, were discussed in theory. The size distribution of the objects segmented in this way indicated that optimisation of a Fisher discriminant could be further used for classifying objects as small artefacts, singlet RBCs, doublets, or triplets etc. of adjoining cells provided optimisation was via a global search. Application of constraints on the relationships between the sizes of singlet and multiplet RBCs led to a number of tests that enabled clusters of cells to be reliably identified and accurate total RBC counts to be made. Development of an application to make such counts could be very useful both in research laboratories and in improving treatment of malaria. Unfortunately, the very small number of pixels belonging to parasite infections mean that it is difficult to segment parasite objects and thus to identify infected RBCs and to determine the parasitemia. Preliminary attempts to do so by similar, unsupervised means using Fischer discriminants, even when applied in a hierarchical manner, though suggestive that it may ultimately be possible to develop such a system remain on the evidence currently available, inconclusive. Appendices give details of material from old texts no longer easily accessible.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item