Comparison of regularization and data augmentation as means of improving malaria classification using convolutional neural network

Abstract

One of the major limiting factors of the application of deep learning in automating malaria diagnosis is the insufficiency of labeled data for network training which leads to overfitting of the model on training data. Network regularization and data augmentation are two major techniques employed during network design and training to mitigate the effect of overfitting. To investigate and compare the effect of these two techniques, this study developed three deep learning models based on convolutional neural network which were trained on 27558 images of equal instances of malaria (P.Falciparum specie) infected and uninfected cells (NIH malaria dataset). The first model which consist of three convolution layers and one hidden layer achieved a validation accuracy of 95%. A second model was developed which employed network regularization, by training the model on fewer neurons to increase its robustness. Drop out regularization of 20%, 30% and 50% were embedded on successive convolution layers which resulted in a final validation accuracy of 95.1%. Data augmentation in the form of various image transformation was employed. This resulted in a validation accuracy of 93.7%.