Malaria detection using deep residual networks with mobile microscopy

Abstract

Automatic segmentation of erythrocytes in microscopic blood smear phone images is a critical step to visualize and identify malaria using machine learning technologies. However, it still remains a challenging problem due to the scarcity of experts, low image qualities, slow manual and inefficient quality of diagnosis. To handle these issues to some extent, we proposed an effective multi-magnification deep residual neural network (MM-ResNet), where we fully automatically classify the microscopic blood smear images as either infected/ non-infected at multiple magnifications. We have experimentally evaluated our approach by using it to train more efficient variants of different compact deep convolutional neural networks (CNN), evaluated on phone datasets. The MM-ResNet end-to-end framework shows similar or superior accuracy than the baseline architectures, as measured by GPU timings on the publicly available microscopic blood smear phone images. This approach is the first application of a MM-ResNet for malaria-infected erythrocyte identification in microscopic blood smear images.