Chest X-rays could be helpful in determining whether or not a person is infected with the novel Coronavirus that causes COVID-19 disease. This diagnosis could even be made computer-aided, using advanced artificial intelligence techniques, such as Deep Learning.
This is the thesis pursued by a group of researchers from universities in the United Kingdom and Spain, led by Eng. Saúl Calderón Ramírez, professor and researcher at the Technological Institute of Costa Rica (TEC), who is conducting doctoral studies at the Institute of Artificial Intelligence, Montfort University, in Leicester, England.
“The topic of interest is to develop an artificial intelligence model to estimate, from chest X-ray images, COVID-19 damage. It would be to use this as an assistance tool for the diagnosis of the disease. Our goal is to develop an artificial intelligence system that hopefully improves diagnostic sensitivity compared to a normal radiology test and that then can be used as a triage or screening system to detect patients who have COVID-19 damage, in a more fast manner,” says Calderón.
The problem, explains Calderón and other researchers in their study, is that “Deep Learning Systems” typically work with large databases to train computers so that they can identify images with a high percentage of success. But it is difficult to get data labeled COVID, due to the rapid advance of the Pandemic.
So the specialists suggest that a viable solution is to implement “semi-supervised” algorithms that help the system to deal with the scarcity of data. This is detailed in a paper accepted for the 25th edition of the International Conference on Pattern Recognition (ICPR), which should have been held in 2020 in Italy, but was postponed to the beginning of next year for reasons of the Pandemic.
“What we did was create a model, using semi-supervised learning, since there are very few images marked or that we know are of COVID-19. So made a model with unmarked images, and obtained very good results”, Calderón states.
With this method, it is possible to reach 90% success in identifying patients with COVID-19, 15% more than without the semi-supervised algorithm, detailed Calderón and the other researchers, from universities in Costa Rica, Spain and the United Kingdom.
According to Calderón, this diagnostic method could be of great help for medical centers in remote communities, where access to COVID-19 tests is limited, such as PCR (Polymeric Chain Reaction), which are the most used.
“Access to X-ray equipment is generally easier or more widespread than other technologies, so it is sometimes easier to do an X-ray than it is to have access to a laboratory for DNA analysis. That is why we believe that a system of this type, like the one we are studying, can be used as a diagnostic method”, highlights the researcher.
These artificial intelligence systems can be accessed remotely, so a technician or doctor could send the X-ray image with their cell phone and receive a response shortly after.
Calderón’s team is not the only one studying how data analysis can be useful to health or medical authorities in the face of the Pandemic. From epidemiological management to patient care, these technologies added to artificial intelligence can be very useful.
An example is the study by researchers from Kings College London, in which they analyzed data of 1,600 people from the United States and the United Kingdom (collected by an App) to define that there are “six types of COVID”, or six typical forms in which that patients are affected.
The use of machine learning algorithms also served to identify the progression of symptoms, so it could be of great use to doctors to identify how the disease could affect a certain patient and how to treat it preventively. This was later confirmed with another database, of 1,000 users from the United States, United Kingdom and Sweden.
Similarly, Calderón affirms that more information could be added to the system they are developing, such as the symptoms that patients present, to improve accuracy levels.
“Building a system where we include not only X-ray image data, but also symptomatic patterns, physiological, metabolic, and other data, could generate a model to make much more precise estimates,” argues Calderón.
In a second advance, Calderón and the other researchers seek to improve the effectiveness of the semi-supervised algorithm and, at the same time, make it work with different data applications.
This would allow the development of “tailor-made” diagnostic tools for different countries, appropriate to the way in which the disease is affecting a given area. Specifically, it seeks to apply the technique with data from Costa Rica.
“A Virus like this has different behaviors, it mutates, there are different strains and it may be expressed differently on X-rays. That is why it is important to build a custom model that responds to the behavior of the disease in a given country. It would not be advisable to use an artificial intelligence model developed with data from the United States to diagnose patients from Costa Rica, because there may be subtle differences with respect to the manifestations of the disease visible on X-rays,” concludes Calderón.