Online training and educational games for malaria diagnosis

Background: Optical microscopy for malaria diagnosis is a time-consuming task that requires trained diagnosticians to manually scan blood-smear slides. In resource-constrained settings, such diagnostic services are often limited by a lack of staff and inadequate training for these workers. With our BioGames platform, we aim to provide distributed biomedical image analysis using crowd-sourcing, telepathology, and a gold-standard image library for the training of diagnosticians. We have developed a web-based training game that teaches players to identify malaria-infected red blood cells. Here, we test whether an untrained crowd of gamers could collectively approach the diagnostic accuracy of trained diagnosticians. Methods: We combined the diagnoses of the untrained crowd using a maximum a posteriori estimator and compared them with the diagnoses of an expert diagnostician. Additionally, we generated an image library by using an expectation-maximisation algorithm to combine the positive (infected), negative (uninfected), and questionable (where the image does not provide sufficient information to make a reliable diagnosis) diagnoses of a group of nine expert diagnosticians. Findings: We showed that the BioGames platform could combine the diagnoses of 989 untrained gamers to reach an accuracy of 98.13% of that of an expert diagnostician, with accuracy, sensitivity, and specificity increasing with the number of crowd diagnoses per cell. Also, we generated a library of 2888 red blood cell images, with approximately 4% labelled positive and approximately 5% labelled as questionable, as collectively diagnosed by nine expert diagnosticians. On April 25, 2014 (World Malaria Day), we released this library as an educational training game for diagnosticians, where players diagnose a randomly selected set of approximately 500 cells in each game and receive a quantified score and feedback in the form of misdiagnosed cell images. Interpretation: Platforms such as BioGames could provide crowd-sourced biomedical image-analysis services. Our method relies on machine-learning algorithms and, therefore, is expected to be even more accurate as the number of gamers and cell images increase. We plan to expand our image library with new microscopic images of thin/thick blood smears and to promote the widespread use of our digital educational tool for malaria diagnosis training. Funding: Non