Statistical Comparison of Different Machine-Learning Approaches for Malaria Parasites Detection in Microscopic Images

Abstract

Malaria is a severe public health problem across the world, particularly in developing countries (≈80% of the cases occur in Africa), putting at special risk the most unprotected groups of society: children and pregnant women. Since it can be caused by 4 different species of parasites, each having different stages of evolution, approaching the right diagnosis without access to costly equipment is complex. Ergo, research has focused on speeding up and lowering the costs of its diagnosis, by resorting to automatic machine classification of microscopic images. Still, most approaches rely on simplistic, single-model classifiers, with a constant absence of a systematic statistical comparison in the literature that supports a particular technique or feature.Hence, this dissertation presents: (i) design and execute a statistical comparison of different ML approaches for the detection of such parasites in microscopic images, (ii) identify which features are more relevant for prediction, and (iii) identify which models and techniques achieve the best results, balancing precision and recall. Given the stated problem, and before approaching it, it was performed an initial statistical analysis to the dataset, to discover its proportions and to detect highly correlated features.After knowing the data, it was developed a framework that (i) optimizes the values of the considered classification and feature selection algorithms, (ii) computes a statistical comparison of different machine learning approaches to the same dataset, using different metrics on the cross validation, where there were used different metrics to measure the performance value variation and evaluate which one is consistent with the data, and, finally, (iii) performs a statistical hypothesis test,to guarantee that the data model with the best performance is distinct from all the others considered in this study. As result, one can verify an improvement over the established baseline, by using a Fdr feature selection method followed by a Ada Boosting classifier with 350 estimators.