Robust deep neural network model for identification of malaria parasites in cell images

Abstract

Malaria is a common mosquito-borne disease that is transmitted through humans by mosquito bites. Severe cases of it lead to death. In the labs, tests are performed on the thick and thin blood smears to determine whether there is any presence of malaria parasites. Counting the number of parasitized and normal cells is a tremendously trivial and important task that relies mostly on the proficiency of the observer. Recent advancement in computer-aided diagnosis systems has enabled it to use sophisticated machine learning and deep learning algorithms. We propose the usage of CNN models for the classification of the cell images as CNN is capable of identifying many complex features and labeling them appropriately. Here we have implemented a variant of CNN called VGG16 model for the identification. In order to improve these models' robustness from adversarial attacks like from fast sign gradient attack, we have trained our model with adversarial images with correct labels to ensure accurate classification even when adversarial images are used for classification. Our model provides 95.96% accuracy before adversarial training and 95.79% accuracy after adversarial training. For adversarial images, we acquire 93.38% accuracy.