Scientists at the Communications Engineering Branch (CEB) of the Lister Hill National Center for Biomedical Communications (LHC) at the National Library of Medicine (NLM) have done research in deep machine learning methods and image analysis for malaria screening and diagnosis. To enable their research, acquisition of big malaria image data was essential. With the support of collaborators in the field, LHC acquired about 3000 thick smear images from 200 patients at Chittagong Hospital in Bangladesh. An expert on site manually annotated 85,000 parasites and more than 60,000 white blood cells in these images, allowing training of deep neural networks. Together with 1300 fully annotated thin smear images, which have been acquired earlier and which have infected and uninfected red blood cells marked for training and testing, this set is one of the largest image repositories in the world for malaria blood smears. LHC is going to make this repository publicly available to the research community, and has already released 27,000 cell images of parasitized and uninfected red blood cells. In machine learning research, LHC has been one of the first institutes to apply deep learning to malaria diagnosis in blood smears. LHC has shown that deep machine learning methods can outperform traditional machine learning methods and can provide high sensitivity and specificity. In fact, LHC has been able to train deep neural network classifiers for both thin and thick smears. Furthermore, research has revolved around finding small and powerful network models that can be executed on smartphones for field use. LHC has successfully implemented a prototype smartphone application for processing thin smears, which achieves the same performance as a full-fledged implementation on more powerful stationary hardware. The search for a similar portable model architecture for thick smears is ongoing, with first experiments showing promising results. Other research into machine learning has focused on using deep learning for detecting and segmenting individual blood slides and parasites to cope with the segmentation problem of touching or overlapping cells. For image analysis, research has concentrated on solving key problems of blood smear imaging. Specifically, LHC investigated methods that can detect white blood cells in thick and thin smears, which is an important part of the malaria diagnosis protocol. LHC has also studied different methods for finding parasite candidates in thick smears, including both traditional methods and deep learning methods. Other research performed in image analysis has produced methods for normalizing or reducing staining variations, enhancing image quality, and improving illumination. To validate the research results and to acquire more training and testing data for research into big data, LHC has made its smartphone software available for download so that interested researchers can collaborate. About twenty experts in five countries are currently collaborating with LHC, contributing images, annotations, and valuable feedback. LHC is planning to make its research available via the Open Microscopy Environment (OMERO), where researchers can upload their images and have them processed by LHCs machine learning and image analysis methods.
| Poostchi, Mahdieh; Silamut, Kamolrat; Maude, Richard J et al. (2018) Image analysis and machine learning for detecting malaria. Transl Res 194:36-55 |
| Rajaraman, Sivaramakrishnan; Antani, Sameer K; Poostchi, Mahdieh et al. (2018) Pre-trained convolutional neural networks as feature extractors toward improved malaria parasite detection in thin blood smear images. PeerJ 6:e4568 |
