Artificial Intelligence Offers a Better Way to Diagnose Malaria
An algorithm for spotting malaria under the microscope could bring accurate, rapid diagnosis to understaffed areas.
For all our efforts to control malaria, diagnosing it in many parts of the world still requires counting malaria parasites under the microscope on a glass slide smeared with blood. Now an artificial intelligence program can do it more reliably than most humans.
That AI comes inside an automated microscope called the Autoscope, which is 90 percent accurate and specific at detecting malaria parasites. Charles Delahunt and colleagues at Intellectual Ventures Laboratory—the research arm of Nathan Myhrvold’s patent licensing company Intellectual Ventures in Seattle—built the system with support from Bill and Melinda Gates through the Global Good Fund. The Autoscope was tested in the field at the Shoklo Malaria Research Unit on the Thailand-Myanmar border during malaria season in December 2014 and January 2015. The results were published in December.
The Autoscope, a 15-inch-tall, 7-inch-wide smooth white box enclosing a microscope with an attached laptop running a software algorithm, uses deep learning to analyze microscope images. Deep-learning software uses artificial neural networks mimicking the brain to allow computers to recognize abstract patterns. Delahunt’s team trained the software on 120 slides from collections around the world, both with and without malaria. The software uses visual features like shape, color, and texture to calculate the probability that a given object is a malaria parasite. It classified 170 samples during the field testing in Thailand.
"It could have broad applicability, not only in research and surveillance of antimalarial drug resistance but also in clinical practice,” says Mehul Dhorda, head of the Asia Regional Centre at the WorldWide Antimalarial Resistance Network. Dhorda works with Intellectual Ventures on some of the current Autoscope trials but was not an author of the research.
In 2015, malaria affected 214 million people and killed an estimated 438,000. Worldwide, we spend about $2.7 billion a year on fighting and controlling malaria.
Current diagnosis of malaria relies on two approaches: microscopy and rapid diagnostic tests. Rapid diagnostic tests are portable cards that display bands in the presence of malaria, much like an at-home pregnancy test. They’re inexpensive, but even a small cost can be prohibitive. By contrast, once a clinic has a microscope and some glass slides, they can be reused indefinitely without further costs.
Another disadvantage of rapid diagnostic tests is that they don’t quantify malaria. They only detect its presence or absence, so they are not ideal for cases of drug-resistant or severe malaria. “If you have a severely ill child with severe malaria, then it's important that you control the parasite density. Every six hours you want to see, is it coming down? Is my treatment having an effect?” says Albert Kilian, a public health and malaria expert at Tropical Health consultancy. “And in these cases, [rapid diagnostic tests] don't serve, so you need to count [the parasites].”
The microscopy currently used to quantify the parasites requires well-trained microscopists, and many malaria-prone areas don’t have enough of those, or the resources to train new ones. By contrast, anyone can use the Autoscope. “We’re not as good as the very best humans, but we’re certainly better than most microscopists in the field,” according to the World Health Organization’s standards, says Delahunt.
There are obstacles to overcome before the Autoscope can be used where it’s needed most. The device requires electricity, so it’s useless in areas that lack adequate power.
And then there is the cost issue. Intellectual Ventures Laboratory is currently looking for a commercial partner to help it drive down the cost of the Autoscope to $1,500 to $4,000. It also plans to spend 2016 testing the Autoscope in more field trials in Peru and Southeast Asia, including some tests for drug-resistant malaria cases.
Hear more about deep learning at EmTech MIT 2017.Register now