AI algorithms for rapid coronavirus diagnosis

An international team of researchers trained an artificial intelligence (AI) model to sort and combine radiology data and clinical information to rapidly and accurately diagnose patients with COVID-19.

“The AI system could rapidly flag suspected cases so radiologists can review those with a higher priority,” says Zahi Fayad, study co-author and professor of radiology and medicine at the Mount Sinai School of Medicine, New York, USA. “If radiologists also diagnose these AI-identified patients as COVID, these patients can be isolated before they get their RT-PCR test results.”

The researchers trained their system using datasets from 905 patients from different medical centres and hospitals across China between January and March 2020. The trained model was tested on a set of 279 patients and performed well, with a confidence rate of 84%.

The model improved detection by correctly identifying COVID-19 infection in 17 out of 25 patients who presented with normal CT scans and who had previously been classified as COVID-19 negative by radiologists.

The algorithms integrate chest CT scans with information such as a patient’s age and sex, symptoms, and exposure history to identify a SARS-CoV-2 infection in the early stages. The model can work alongside routine tests, such as reverse transcriptase polymerase chain reaction (RT–PCR), to reduce the frequency of false negatives and decrease the risk of viral transmission.

Fayad explains that they created three models for testing: a deep convolutional neural network (CNN) that only used CT scans; a model that used support vector machine (SVM) and random forest and multilayer perceptron (MLP) to classify clinical data; and a joint one that integrated CT scans with clinical data.

Patient data were entered into the AI system as numeric values of 1 and 0—with 1 representing the presence of symptoms, confirmed travel to Wuhan or close contact with positive cases, and 0 representing no symptoms or link to Wuhan or anyone carrying the virus. Chest CTs were input into the system as pixel values. The data is then converted into vectors that are combined and used for making predictions.

The proposed AI system is fast compared to conventional diagnostic methods, yielding results in a minute or less using GPU-powered computers, and in a few minutes with CPU-based machines. RT-PCR tests typically take two days to complete.

For the system to be reliable, however, researchers say that they would need to collect more scans and data from multiple countries. The proposed model also carries a few limitations, such as a bias in the training data towards patients with COVID-19, as opposed to other infections and pneumonias that exhibit similar symptoms.

The researchers plan to improve predictions by updating their model architecture to a three-dimensional mesh that incorporates more frames from CT imaging. Fayad explains that “the current model was developed at slice level. Only key suspected frames were used from CT scans due to the trade-off between efficiency and turn-around time.”

References