CheXNeXt: an AI checks chest X-rays for 14 findings at once and goes up against 9 radiologists

Why do chest X-rays matter so much?

The chest X-ray is the most common imaging test in the world, with over 2 billion done each year, according to the authors. It helps doctors find common conditions such as pneumonia, tuberculosis, and lung cancer.

But reading X-rays takes time and requires a radiologist, a doctor who specializes in medical images. Many parts of the world do not have enough of them, and even experienced radiologists make more mistakes when they are tired. The researchers wanted to know if an AI could read these images as well as practicing radiologists.

What did the Stanford team do?

The team trained CheXNeXt, a convolutional neural network (a type of AI built to read images), on a large public dataset: 112,120 front-view chest X-rays from 30,805 patients. Each image had up to 14 labels, such as pneumonia, pleural effusion (fluid around the lung), atelectasis (a partly collapsed lung), nodules, or masses.

For the test, they picked 420 X-rays the AI had never seen. The correct answer for each image was set by a vote of 3 radiologists who specialize in the chest. The AI was then compared with 9 radiologists: 6 board-certified radiologists and 3 senior radiology residents.

How did the AI compare with radiologists?

Results are given as the area under the curve, a score from 0.5 (chance) to 1 (a perfect test). CheXNeXt performed as well as the radiologists on 10 findings and better on one, atelectasis (0.862 vs. 0.808).

It did worse on 3 findings: cardiomegaly (an enlarged heart), emphysema (damaged, overinflated lungs), and hiatal hernia (part of the stomach pushing up into the chest). All three were uncommon in the training images.

For masses and nodules, possible signs of lung cancer, the AI caught a larger share than the board-certified radiologists. It read all 420 X-rays in 1.5 minutes, vs. 240 minutes on average for the radiologists, and it produced heat maps showing which part of the image caught its attention.

What could this mean for patients?

The authors say a tool like this could help sort the worklist, so the sickest patients' X-rays get read first. It could also help in places with no radiologist on hand, for example to look for signs of tuberculosis or lung nodules.

The radiologist would still make the final call. For patients, the benefit would be faster results, especially where specialists are scarce.

What are the limits of this study?

This was a retrospective study: it used older images, all from a single institution, so there is no guarantee the AI would do as well elsewhere. Both the AI and the radiologists saw only the front-view X-ray, without the patient's history or earlier exams, which radiologists normally use.

So the test does not reflect real practice. The authors say the algorithm now needs to be tested prospectively, in real hospitals, before it is used in patient care.