Classification of Plasmodium-Infected Erythrocytes Through Digital Image Processing

Abstract

The development of antimalarial drugs requires performing laboratory experiments that include the analysis of blood smears infected with Plasmodium. Analyzing visually the resulting microscopy images is usually a slow and tedious task prone to errors due to fatigue and subjectivity of the analysts. These facts have motivated the creation of digital image processing systems to automate this analysis. In this work a computer vision solution to process microscopy images of blood smears containing erythrocytes infected with Plasmodium is shown. This system performs tasks like illumination and color correction, image segmentation including splitting of clumped objects and extraction of color features. A set of different classifiers was tested and evaluated to find the best one in terms of indexes of effectiveness. A new feature named pixels fraction was introduced and used together with a number of other color-related features, from which a subset to classify cells into normal or infected was selected. The classifiers evaluated were: support vector machines (SVM), K-nearest neighbors (KNN), J48, Random Forest (RF), Naïve Bayes and linear discriminant analysis (LDA). All of them were evaluated in terms of correct classification rate, sensitivity, specificity, F-measure and area under Receiver Operating Characteristic (ROC) curve (AUC). The effectiveness of the pixels fraction as a new feature was demonstrated by the experimental results. In regard to classifiers, J48 and Random Forest showed the best results.