AI boosts CT's COVID-19 diagnostic performance

2020 04 22 18 18 5522 Coronavirus Cells Lungs 400

Artificial intelligence (AI) boosts the performance of chest computed tomography (CT) when combined with clinical history and laboratory data in diagnosing COVID-19 pneumonia, according to a study published May 19 in Nature Medicine.

Chest CT has been shown to be a sensitive and specific modality for diagnosing COVID-19, wrote a team led by Yang Yang, PhD, of Icahn School of Medicine in New York City. Yet CT results can come back negative in patients with early-stage disease. That's why AI could make CT an even more effective tool in the COVID-19 diagnostic arsenal.

"Chest CT findings are normal in some patients early in the disease course and therefore chest CT alone has limited negative predictive value to fully exclude infection," the group wrote. "We propose that AI algorithms may [address this problem] by integrating chest CT findings with clinical symptoms, exposure history, and laboratory testing in the algorithm."

Quick diagnosis of COVID-19 is key not only to effective patient treatment but also to curbing the disease's spread, the group noted. The current method, reverse transcription polymerase chain reaction (RT-PCR) testing, can take up to two days to complete, and it may require repeat testing.

Chest CT can be an effective tool for diagnosing COVID-19, though it's not perfect: It can't necessarily distinguish COVID-19 from other lung disease. In fact, its use has been debated, and the American College of Radiology cautions that it should be used sparingly.

Yang and colleagues sought to investigate how AI could help make chest CT for diagnosing COVID-19 more effective. The study included CT scans and clinical information from 905 patients treated at 18 medical centers in China. Of the 905 patients, 419 (46.3%) had positive RT-PCR results.

The investigators trained the AI model on a test set of 279 cases taken from the 905 CT exams, then compared its performance with that of a senior thoracic radiologist and a fellow. The researchers found that the AI system had an area under the curve (AUC) of 0.92 and sensitivity comparable to the senior thoracic radiologist.

Performance comparison of AI algorithm and radiologists for identifying COVID-19 on CT
Performance measure Thoracic radiology fellow Senior thoracic radiologist AI algorithm
AUC 0.73 0.84 0.92
Sensitivity 56% 74.6% 84.3%
Specificity 90.3% 93.8% 82.8%

The AI model also improved the detection of patients who tested positive via RT-PCR but had apparently normal CT scans, identifying 17 of 25 patients as having COVID-19 (68%), while the radiologists classified all 25 as negative for the disease.

The study findings show that when CT scans and associated clinical history are available, AI can boost CT's COVID-19 diagnostic performance, according to the team.

"The AI system could be implemented as a rapid diagnostic tool to flag patients with suspected COVID-19 infection when CT images and/or clinical information are available and radiologists could review these suspected cases identified by AI with a higher priority," it concluded.

Page 1 of 120
Next Page