At the Radboudumc in Nijmegen, several scientists have been engaged in internationally leading research on COVID-19 since March. Among those, an AI software program has been developed with which, based on a CT scan or X-ray, it can be estimated whether someone has covid.
During the first corona wave, there was a great demand for software that could predict whether someone had covid. “People who feared to have been infected reported to the hospital doors, but hardly any tests were available,” says Bram van Ginneken, professor of medical image analysis at the Radboudumc and founder of the software company Thirona, a spin-off of the Radboudumc that focuses on AI applications for medical images.
At the beginning of the corona crisis, physicians were almost desperate for opportunities to quickly assess whether a patient was infected with covid-19. Van Ginneken: “The Radboudumc came up with an urgent demand think along with them. Our researchers have been working for some time with algorithms that enable quick analysis of medical lung images, for example for lung cancer or TB. Like everyone else, we worked from home, but we all focused on developing software to quickly and accurately assess CT scans for COVID-19.”
Algorithm available for free
Within just a few months, Corads-AI – the algorithm that can estimate whether someone has COVID-19 – was ready for the market. Corads-AI is described in the leading journal Radiology and is made available free of charge during the pandemic via grandchallenge.org. “Corads-AI is widely used worldwide, particularly in North and South America, Europe, Asia, and Russia. Doctors can add their own CT scan and have it reviewed by the program.”
AI software performs better than a rapid test
Now that more tests are available worldwide and also many rapid tests are available, a CT scan assessment is not the most obvious instrument to determine COVID-19 in people with complaints. “But microbiological tests don’t always do better,” says Van Ginneken. “If you put a stick through someone’s nose and throat, the virus has to be present exactly where you scrape. It sometimes happens that the virus has already descended to the lungs and someone tests negative, but the virus is actually present in the lungs. On a CT scan, you usually see this immediately, even at an early stage.”
Although the course of the disease varies per patient, a CT scan of covid lungs is often surprisingly recognizable. “Pneumonia due to COVID-19 usually has a typical pattern with abnormalities at the edges of both lungs that is fairly easy to recognize with our algorithms.”
X-rays in Africa
Corads-AI is widely used in Europe and elsewhere, but in Africa, a CT scan is a luxury that many hospitals cannot afford. Van Ginneken has already developed AI software to determine whether someone has TB by means of an X-ray. “In Africa, patients come to a hospital coughing, and you can’t tell whether someone has TB, COVID-19, or any other condition. It is very important to differentiate between these syndromes. TB has a more serious course for young people, but it is also easy to treat with antibiotics. You want to be able to separate patients properly. Here, too, a newly developed AI program offers a solution: CAD4COVID. Van Ginneken: “Analyzing the type of lung damage gives a reasonably good answer to what the patient’s problem really is. In about 80 percent of situations, the algorithm can correctly indicate which patient has covid. If this is supplemented with a simple blood test, it rises to about 90 percent.”
In the small African country of Lesotho (which lies within the borders of South Africa) an operational trial was started with this combination of CAD4COVID and blood testing. Although there are now more tests to determine covid and there may soon be a vaccine, it remains to be seen to what extent they will be able to use it in developing countries in the short term. “There are currently PCR test devices available for Africa, but they cannot get the cartridges needed for the assessment on covid-19.”
Corads-AI to become superfluous?
Of course, all eyes worldwide are on the advent of the vaccine. Van Ginneken: “We all hope for a well-functioning vaccine, but covid will most likely continue to exist and in the future, when assessing a CT scan, you want to determine whether someone has or has had covid.”
Currently, Van Ginneken and his colleagues are also working on new research questions surrounding covid-19 that can also be predicted by algorithms. Which patients with which lung image can best be treated? Van Ginneken: “For some patients, is it wise to start with dexamethasone, others can be helped with remdesivir. With AI you can make statements about this. The problem is that we don’t yet have a good database in the Netherlands, which is necessary for our software to be able to learn. The demand is also changing all the time because the developments surrounding treatments are going very fast.”