What contribution can Artificial Intelligence make against COVID-19 disease?
Prof. Arturo Chiti, Head of Nuclear Medicine Unit – Istituto Clinico Humanitas
Recognizing interstitial pneumonia with a CT scan or chest X-ray. Cross-checking clinical information and epidemiological data. Evaluating the course of a disease and assessing the risk for the patient. Choosing the best treatment strategy. Specialists do this job successfully every day in hospitals and clinics. But what happens during a pandemic crisis, when the number of patients becomes suddenly very large, the amount of data is huge and resources are limited? What tools are offered to doctors by new technologies to deal with the emergency?
In recent times there has been a lot of talk about Artificial Intelligence and the help it could provide in medicine. In situations such as the crisis caused by the global spread of the Sars-CoV-2 virus, AI can make a decisive contribution providing a formidable instrument, potentially available to all doctors. Artificial Intelligence applications have been spreading in clinical practice for a few years now. Diagnostic imaging is one of the areas in which AI has had a great development. Software are available which help radiologists and nuclear doctors to evaluate the images. Systems that help to identify lesions or recognize alterations, for example in the oncological or neurological field. Other more sophisticated tools are able to give a meaning to the alterations, or to optimize the diagnostic method or flows. The applications are therefore multiple. In the coming years, they will have a development in the direction of an increasing level of automation.
What contribution can Artificial Intelligence make against COVID-19 disease? As we said, applications are already available that allow to recognize interstitial pneumonia, which is the typical clinical picture of patients infected with Sars-CoV-2. Commercial software has been developed in China, others may be requested on free loan or made available online. We are talking about very advanced software that are able to recognize structures and alterations.
How do they learn? The software is trained by experienced radiologists: they indicate the characteristics of the structure (for example pneumonia in an x-ray plate) and the algorithm, after analyzing thousands of images, it begins to detect and recognize them. It is a challenging job, but of great value. At the end of the training, the application will be able to complete the task alone and can be used on any computer in the world.
The more complex is the task of those who have to read the images, the more important is the support provided by the software. The tool can be of great help in the less obvious or “borderline” cases, for example in the presence of pneumonia in the initial phase.
With artificial intelligence algorithms, the knowledge of more expert professionals is transferred into a software, allowing everyone to make reliable diagnoses. In the current case, the impact can be high especially in those hospital where the infection has not yet arrived massively and radiologists have had less opportunity to gain experience.
Does this mean that Artificial Intelligence, left alone, could diagnose COVID-19 disease? It’s not that simple. The result of imaging diagnostics is not enough to recognize Sars-CoV-2 pneumonia. The radiological picture is very similar to other types of pneumonia. However, there are clues that can make us lean towards one diagnosis over another. It needs to be assessed whether a pandemic is ongoing. If the peak influence period has passed. If the patient has had contact with infected people. We do not have a diagnosis without considering the clinical and epidemiological context.
Here AI can play an important role, thanks to its ability to process data from different sources. There is one more step. Can Artificial Intelligence make predictions about prognosis and understand the evolution of the disease? This is what researchers from Humanitas intend to verify. As said, the prediction of the clinical course cannot be made only on through images: clinical and epidemiological data will also have to be considered.
Neural networks are extremely effective in making classification. The software can evaluate the CT scan, but also age and provenance of the patient, co- morbidities, blood test results and oxygenation, and therefore place the patient in a risk class. Doing so, it could help to decide whether a patient should be hospitalized or treated at home.