Deep learning enabled multi-wavelength spatial coherence microscope for the classification of malaria-infected stages with limited labelled data size

Malaria is a life-threatening mosquito-borne blood disease and therefore, early detection is essential for health. The standard detection technique is a Giemsa-stained blood smear test, which requires a highly skilled technician. Automated classifications of different phases of malaria continue to be a challenge with limited label data, especially in the case of early trophozoite and late trophozoite or schizont phase where the sensitivity is lower. The aim of the study is to develop a fast, robust and fully automated malaria classification system with limited labeled data set by using a multi-wavelength based system and a pre-trained convolutional neural networks (CNNs). The multi-wavelength system improves the classification performance by increasing the training dataset. We also compare our customized CNN performance with other conventional CNNs and show that our network with a lesser number of layers has a comparable performance. We believe that our approach can be applied to other limited annotated biological datasets.