Abstract—Malaria prognosis, performed through the identi-fication of parasites using microscopy, is a vital step in theearly initiation of treatment. Malaria inducing parasites suchas Plasmodium falciparum are difficult to identify and thus havea high mortality rate. For these reasons, a deep convolutionalneural network algorithm is proposed in this paper to aid inaccurately identifying parasitic cells from red blood smears. Byusing a mixture of machine learning techniques such as transferlearning, a cyclical and constant learning rate, and ensemblemethods, we have developed a model capable of accuratelyidentifying parasitic cells within red blood smears. 14 networkspretrained from the ImageNet database are retrained with thefully connected layers replaced. A cyclical and constant learningrate are used to traverse local minima in each network. Theoutput of each trained neural network is representing a singlevote that is used in the classification process. Majority votingcriteria are applied in the final classification decision betweenthe candidate malaria cells. Several experiments were conductedto evaluate the performance of the proposed model. The NIHMalaria Dataset from the National Institute of Health, a datasetof 27,558 images formed from microscopic patches of red bloodsmears, is used in these experiments. The dataset is segmentedinto 80% training set, 10% validation set, and 10% test set. Thevalidation set is used as the decision metric for choosing ensemblenetwork architectures and the test set is used as the evaluationmetric for each model. Different ensemble network architecturesare experimented with and promising performance is observedon the test dataset with the best models achieving a test accuracybetter than several state-of-the-art methodologies.
Published: October 2020
Read More