X-rays allow rapid diagnosis of people infected with COVID-19 virus. The images are reviewed by radiologists. Artificial intelligence (AI) can support them in this process. For instance, AI can predict whether someone has infiltrative lung abnormalities as well as radiologists.
This is according to an initial study by the Jeroen Bosch Hospital, the results of which were published in the scientific journal Radiology. It was previously shown that AI is able to diagnose tuberculosis as accurately from an X-ray as a radiologist.
Shortage of test kits and expertise
The Jeroen Bosch Hospital sees several opportunities to use AI during the corona crisis and the scarcity of test kits to test for COVID-19. "Radiography is a fast and relatively cheap imaging modality available in many healthcare settings where resources are scarce. Unfortunately, there is a severe shortage of radiological expertise in this region for accurate interpretation of such images. An AI system can be a useful tool to complement radiologists or in the common case when radiological expertise is not available," write researchers Keelin Murphy, Henk Smits, Arnoud J.G. Knoops, Mike B.J.M. Korst, Tijs Samson, Ernst T. Scholten, Steven Schalekamp, Cornelia M. Schaefer-Prokop, Rick H. H. M. Philipsen, Annet Meijers, Jaime Melendez, Bram van Ginneken and Matthieu Rutten in an article in Radiology.
"Previous work related to the detection of tuberculosis using radiography shows that software can perform at the level of a radiological expert in identifying tuberculosis. In this study, we evaluate the performance of an available AI system for the detection of COVID-19 pneumonia."
Deep learning
The study used CAD4COVID-XRay, an AI system that uses deep learning. This is a form of AI where an algorithm is trained by having it search for patterns in a huge dataset. A deep learning algorithm continues to learn continuously and thus becomes smarter and smarter over time.
CAD4COVID was developed by Delft Imaging and Thirona. The software analyses chest X-rays of patients with suspected COVID-19 infection. The software then gives a score between 0 and 100, shows the COVID-19-related abnormalities in the lungs and reveals which part of the lungs is affected.
Training algorithm
CAD4COVID-XRay was initially trained by Jeroen Bosch Hospital using a dataset of 22,184 lung images, 7851 of healthy lungs and 5,012 of pneumonia. This since comprehensive datasets of pulmonary photos with pneumonia caused by COVID-19 are difficult to obtain. This initial training taught the algorithm to recognise pneumonia. The algorithm was then further trained with a smaller dataset of 416 lung photos with pneumonia caused by COVID-19 to specifically identify COVID-19.
Following this training, 454 X-ray images from the Jeroen Bosch Hospital were reviewed by both the AI system and about six radiologists. 231 of these images are from patients who tested negative for the virus and 223 from patients who tested positive. During the experiment, the AI system managed to assess all X-ray images as well as the radiologists.
Further training
Following the study, the hospital wants the deep learning algorithm to learn further using more images. These may be combined with laboratory results.
The full study can be found here.
Author: Wouter Hoeffnagel
