Northwestern AI algorithm outperforms radiologists in COVID-19 chest X-ray detection

In the early months of the pandemic, as Chicago's hospitals filled with COVID patients, researchers at Northwestern University began asking themselves a practical question: could the artificial intelligence tools they had been building for cardiac imaging be adapted to fight a virus spreading through the lungs?

The answer came in the form of DeepCOVID-XR, a machine learning algorithm that learned to read chest X-rays with a speed and accuracy that outpaced the specialists trained to do the same work. The system analyzed a set of 300 test images in 18 minutes. Five experienced chest radiologists, each with specialized training in cardiothoracic imaging, took between two and a half and three and a half hours to examine the same images. When the results were tallied, the radiologists achieved accuracy rates between 76 and 81 percent. The algorithm scored 82 percent.

The difference matters not because machines are inherently superior to human expertise, but because in a pandemic, time collapses into a critical variable. A patient arriving at the hospital for an unrelated reason—a broken bone, chest pain, a routine checkup—might carry the virus without knowing it. A chest X-ray is already routine, inexpensive, and safe. If an algorithm could flag that patient within seconds, that person could be isolated before infecting healthcare workers or other patients. The actual confirmation would still require a COVID test, which could take hours or even days to return results. But the isolation could begin immediately.

To build and train DeepCOVID-XR, the Northwestern team used 17,002 chest X-rays—the largest dataset of COVID-era imaging ever assembled for this purpose. Of those, 5,445 came from patients confirmed positive across Northwestern Memorial Healthcare System hospitals. Aggelos Katsaggelos, the electrical engineering professor who led the work, and Ramsey Wehbe, a cardiologist and postdoctoral AI fellow, had been collaborating on medical imaging projects when the pandemic began. They recognized that the techniques they had developed for reading heart images might translate to reading lungs ravaged by COVID-19.

The virus leaves a distinctive mark on the lungs of many patients. Instead of the clean, orderly appearance of healthy tissue, COVID-positive lungs show irregular, cloudy patterns. Fluid accumulates and inflammation spreads, particularly along the lower lobes and the outer edges. But this visual signature is not unique to COVID-19. Pneumonia, heart failure, and other lung diseases can look similar on an X-ray. Distinguishing between them requires the trained eye of a specialist—and specialists are expensive, not always available, and can only examine so many images in a day.

The researchers were careful to frame their work not as a replacement for testing but as a triage tool. Wehbe emphasized that many COVID patients show no signs of disease on chest X-rays, especially early in infection. In those cases, neither the algorithm nor a radiologist would catch the virus. The system has limits. But those limits are the same limits that human radiologists face. What the algorithm offers is speed and consistency—the ability to process hundreds of images in the time it takes a specialist to process dozens, flagging patients who might otherwise slip through without isolation while waiting for test results to confirm what the X-ray suggested.

The researchers released DeepCOVID-XR publicly, hoping others would continue training it with new data. As of the time of their publication in Radiology, the algorithm remained in the research phase, not yet deployed in clinical settings. But the path forward was clear: in a crisis where time and resources are both scarce, a tool that multiplies the reach of human expertise—even by a small margin—can shift the calculus of who gets protected and when.