Convolutional Neural Networks for Malaria Detection

Abstract

Together with doctors at Haukeland University Hospital in Bergen, we wanted to research how the diagnosis of malaria can be improved. We propose a method that can detect malaria parasites (Plasmodium falciparum) in microscope images. It is based on a convolutional neural network that is trained on over 40000 artificial images. The model performance was evaluated on over 6000 real images of blood smears from Haukeland University Hospital. In the evaluation of the proposed method, we observed that the classification results were different for different classes of microscope images. For most of the classes, the classification was accurate to 85% (or higher), while for some other smaller classes, the accuracy is lover, about 65%. We compared the model against the classification by the trained medical personnel at Haukeland University Hospital. The different comparisons indicate that our model works in a satisfactory manner, and we conclude with some discussion on ideas on which the model can be further improved.