Deep learning approach for detection of Dengue fever from the microscopic images of blood smear

Abstract

Dengue virus (DENV), known to cause dengue fever is a global public health concern. A safe and effective anti-viral drug or vaccine that can protect humans from dengue fever currently does not exist. Today, severe dengue has become a leading cause of serious illness in most Asian and Latin American countries. This digital pathology-related research focuses on the automatic detection of dengue by utilizing digital microscopic peripheral blood smears (PBS). This paper explored pre-trained convolution neural network (CNN) architectures for automatic dengue fever detection. Transfer learning (TL) was performed on two widely used pre-trained CNNs - SqueezeNet and GoogleNet, and employed to differentiate the dengue-infected and normal blood smears. The last few layers were replaced and retrained to customize the architectures for this task. Leishman’s stained dengue-infected and normal control 100x magnified PBS images were included in the study. The best performance was rendered by GoogleNet (Learn Rate, 0.0001; Batch Size, 8) with an Accuracy 91.30%, Sensitivity 84.62%, Specificity 100%, Precision 100%, and F1 score 91.67%. Promising results show that this approach can be an essential adjunct to other clinical methods, namely CBC test & NS1 antigen capture, and can significantly support dengue diagnosis in low-resource setups.