The Lister Hill National Center for Biomedical Communications at NLM is developing a software called MalariaScreener, which can count parasite-infected and uninfected red blood cells, using automatic image analysis and machine learning algorithms. The software, which received an HHS Ventures Award, has been ported from MATLAB to Android smartphones by imaging scientists at NLM and University of Missouri. Running on a camera-equipped smartphone that is attached to a microscopes eyepiece by an adapter, the software screens the field of view for malaria parasites and reports the level of parasitemia to the microscopist. MalariaScreener first identifies red blood cells using segmentation methods such as watershed and level-sets. Then, it computes features that can detect the typical color and shape of parasites. A support vector machine performs the final classification into infected and uninfected cells based on the features computed. MalariaScreener has been trained with more than 200,000 blood cell images acquired from 150 malaria-infected and 50 uninfected patients, which have been annotated by an expert microscopist using a tailored online annotation tool. The encouraging performance of the MalariaScreener prototype in the lab has motivated preparations for testing the system at multiple sites in the field. After field-testing, the smartphone application will be made publicly available for download to other malaria screening sites in the world. With the feedback from expert microscopists, research will continue with implementing sophisticated cell segmentation techniques and deep learning methods to improve the system performance where needed. Another direction of future research is the automatic discrimination between different parasite species and their stage of development. This year we expanded use of Deep Learning for classifying thin blood film cell images and have started implementing a cell classifier for smartphones based on deep learning. Several improvements have been made to our Android smartphone app, including improvements to the interface based on expert feedback and run-time improvements for the machine learning algorithms. For easier remote communication and diagnosis in telehealth applications, the app can now share captured images in a shared cloud folder. Furthermore, multilingual support for Chinese has been added as well as Bluetooth support for picture taking. We have also started acquiring images of thick blood films, including manual annotations of parasites and white blood cells for machine training on this type of images. For further field testing of the app, we have reached out to sites in Thailand and Bangladesh, among others. Investigators interested in testing the app can download the app from the Google Play Store, where we have made it available for the sites participating in the testing.
| 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 |
