Alternative Means to Diagnose COVID-19 Pneumonia

At a Glance

COVID-19 presents with non-specific symptoms that are very similar to other viral illnesses, making it difficult to clinically diagnose. Early in the pandemic, polymerase chain reaction (RT-PCR) testing and x-ray computed tomography (CT) were the primary methods of diagnosis, but they lacked effectiveness. The goal of this project was to develop and deploy an artificial intelligence (AI) solution to assist physicians in achieving rapid and efficient diagnosis of COVID-19 using chest x-ray radiography (CXR).

Researchers were successful in curating a large COVID CXR dataset and ultimately developed an artificial intelligence (AI) solution that could differentiate between COVID-19 pneumonia and non-COVID-19 pneumonia with high sensitivity and specificity. In the future, this dataset will be used to address key challenges in AI including generalizability, interpretability, and algorithmic bias.

The Challenge

The outbreak of COVID-19 began with the initial diagnosis of an unknown respiratory disease presenting with viral pneumonia-like symptoms in late 2019. Patients with COVID-19 present with non-specific symptoms that are very similar to other viral illnesses which make it difficult to establish a clinical diagnosis for the disease. At the time, the reverse transcriptase polymerase chain reaction (RT-PCR) test and x-ray computed tomography (CT) were the primary diagnostic approaches. However, the scarcity of available test kits as well as the high exposure risk, high labor requirements, and low diagnostic specificity of CT imaging hindered the effectiveness of these methods.

Instead, it was recommended that chest x-ray radiography (CXR) images be used to diagnose COVID-19. The major challenge with the use of chest x-rays in COVID-19 diagnosis is its low sensitivity and specificity in radiological diagnosis. The poor diagnostic performance was attributed to the fact that radiologists had no training to differentiate COVID-19 pneumonia from non-COVID-19 pneumonia as many radiologists were seeing COVID-induced pneumonia for the first time as the pandemic unfolded.

Project Goals

The overarching goal of this project was to develop and deploy an artificial intelligence (AI) solution to assist physicians in achieving rapid and efficient diagnosis of COVID-19 using CXR. This goal was approached through three tasks. First, the research team needed to develop and deploy the AI model and make it publicly accessible. Second, the model was evaluated and refined as new data are received. Third, researchers aimed to assist radiologists in identifying COVID-19-specific image features from CXR images.

Results

The research team was successful in achieving the project goal. The team curated one of the world’s largest COVID CXR research datasets by collecting over 50,000 CXR images and their corresponding RT-PCR test results from multiple healthcare institutions. The team used this dataset to develop an AI model that could differentiate COVID-19 pneumonia from non-COVID-19 pneumonia with 88 percent sensitivity and 79 percent specificity. To better understand the AI model’s prediction, the team developed techniques to generate class activation maps which can be used by radiologists to quickly identify potential COVID features in CXR images.

Looking to the Future

The curated dataset developed for this project has important value in the general research of AI. In the future, this dataset will be used to address key challenges in AI such as generalizability, interpretability, and algorithmic bias.