Classical Machine Learning Algorithms and Shallower Convolutional Neural Networks Towards Computationally Efficient and Accurate Classification of Malaria Parasites

Abstract

Malaria and its diagnosis methods need significant attention in Ethiopia. Studies show that for 46 laboratory professionals who were given 6 different positive and negative malaria slides, the detection error rates for plasmodium falciparum and vivax were 43.5% and 37%, respectively. Another similar study reports that the overall malaria diagnosis error rate is 40.4%. To circumvent these challenges, there needs to be a system that automatically and instantly analyzes and manipulates data with less bias so as to reduce misdiagnosis error, cost, workload of physicians and experts, and thereby improving the livelihood of the society. In this study, classical machine learning algorithms and shallow Convolutional Neural Networks (CNN) based classification for parasitized and uninfected malaria images are applied. While classical machine learning algorithms are requiring a feature design step for classification, CNNs learn to recognize objects end to end, implicitly without the need for feature engineering. Many machine learning algorithms such as K-Nearest Neighbors, Random Forest, Gradient Boosting, CNNs etc. are compared in this contribution. From our experimental observation, some of the classical models such as Gradient Boost, Random Forest and Stacked ensemble using color, texture and shape features achieved a classification accuracy of 96%, precision of 97%, recall of 97% and f1-score of 97% for the parasitized samples. A model ensemble of two custom CNNs and mini-VGGNet achieved a precision of 99%, recall of 96% and f1-score of 97% for the parasitized. These results of the CNN models are reproducible for each of them. All the approaches have comparable results, however, the classical machine learning models are resource efficient.