Lung cancer: an AI that reads screening CT scans in 3D, at radiologist level

Why is lung cancer screening hard?

In the United States, lung cancer is the leading cause of cancer death, with an estimated 160,000 deaths in 2018, according to the authors. Screening means giving people at risk a low-dose CT scan, an X-ray examination that exposes the body to less radiation. It has been shown to reduce deaths by 20 to 43%, and it is part of US screening guidelines.

But reading these scans is not simple. Two radiologists do not always agree. And mistakes happen in both directions: false alarms (a false positive, which leads to needless tests and worry) and missed cancers (a false negative).

Better tools matter here because screening only helps if the scans are read well. Too many false alarms tire patients and services, and missed cancers defeat the purpose of screening.

What did the researchers do?

The team built a neural network, a type of AI loosely inspired by the brain, that looks at the scan in three dimensions, as a whole volume rather than slice by slice. It uses the patient's current scan and, when there is one, an older scan, which shows whether a small spot has changed.

They trained and tested it on 6,716 cases from a large US screening trial, then checked it on an independent set of 1,139 cases. Finally, they ran two studies in which six radiologists read scans, to compare their performance with the AI's.

What do the results show?

The AI reached an area under the curve of 94.4% (100% would be a perfect test) on the 6,716 cases, and similar results on the 1,139 independent cases.

When the radiologists had no earlier scan to look at, the AI did better than each of the six: 11% fewer false positives and 5% fewer false negatives (in absolute terms). When an earlier scan was available, the AI performed on a par with the same radiologists.

In plain terms, the AI was at least as careful as experienced readers, and in the harder case without an earlier scan it made fewer mistakes of both kinds.

What could this mean for patients?

A tool like this could act as a second reader: fewer people worried for nothing, and fewer cancers going unnoticed. It could also help centres that lack experienced radiologists and make readings more consistent from one centre to another.

The authors point out that the vast majority of people who could benefit from screening are not screened. Automation could help extend screening around the world while keeping results reliable.

What are the limits of this study?

The AI was evaluated on scans that had already been taken, in test conditions. The study does not show that using this tool with patients improves their health or survival. The radiologists read scans in a study setting, which does not always reflect their daily work.

Prospective trials in real screening centres are needed to see how the AI fits into doctors' work and whether it brings a real benefit.

The model was built and tested on data mainly from a US screening trial, so it will also need checking in other populations and with other scanners.