Accuracy of 82 percent achieved with DeepCOVID-XR compared with 81 percent for consensus of five thoracic radiologists
WEDNESDAY, Nov. 25, 2020 (HealthDay News) — An artificial intelligence (AI) algorithm can detect COVID-19 on chest X-rays with similar performance to that of a consensus of thoracic radiologists, according to a study published online Nov. 24 in Radiology.
Ramsey M. Wehbe, M.D., from the Bluhm Cardiovascular Institute at Northwestern Memorial Hospital in Chicago, and colleagues presented a deep learning AI algorithm (DeepCOVID-XR) for detecting COVID-19 on frontal chest X-rays. The algorithm was trained and validated on 14,788 images (4,253 COVID-19-positive) and was tested on 2,214 images (1,192 COVID-19-positive). Algorithm performance was compared to interpretations from five experienced thoracic radiologists for 300 random test images.
The researchers found that the accuracy of DeepCOVID-XR accuracy was 83 percent, with an area under the receiver operating characteristic curve (AUC) of 0.90 on the entire test set. On 300 random test images (134 COVID-19-positive), accuracy was 82 percent for DeepCOVID-XR compared with 76 to 81 percent for individual radiologists and 81 percent for the consensus of all five radiologists. Sensitivity was significantly higher for DeepCOVID-XR than one radiologist (71 versus 60 percent), and specificity was higher than two radiologists (92 versus 75 and 84 percent). The AUC was 0.88 for DeepCOVID-XR versus the consensus AUC of 0.85; the comparison was not statistically significantly different (P = 0.13 for comparison). Using the consensus interpretation as the reference standard, the AUC for DeepCOVID-XR was 0.95.
“X-rays are inexpensive and already a common element of routine care. This could potentially save money and time — especially because timing is so critical when working with COVID-19,” Aggelos Katsaggelos, Ph.D., senior author of the study, said in a statement.
Copyright © 2020 HealthDay. All rights reserved.