Using artificial intelligence tools to read thousands of chest X-rays and prioritise abnormal ones does not lead to faster lung cancer diagnoses, according to a new study from UCL, UCLH and the University of Nottingham.
The LungIMPACT trial, the largest randomised trial of its kind, ran across five NHS trusts in England and found no significant reduction in the time from X-ray to cancer diagnosis when AI flagged abnormal scans so that a radiologist could prioritise these for early review.
AI prioritisation did speed up one part of the process: the median time for a radiologist to report on a chest X-ray fell from 47 hours to 34 hours. But this improvement did not cascade into faster progress through the rest of the diagnostic pathway. Referral rates, treatment start times, and cancer stage at diagnosis were all comparable across both groups.
“The hypothesis was that getting the X-ray reported faster would cascade into a faster diagnosis, but the NHS system simply could not respond at the same pace,” said UCLH consultant radiographer Nick Woznitza, principal investigator of the trial.
“The bottleneck isn’t the reporting; it’s everything that happens next: telling the patient, the CT appointment, the clinic slot, the multidisciplinary meeting. We’ve shown that AI prioritisation, by itself, cannot fix that,” he added.
Redesign of whole pathway needed
Among the 558 patients diagnosed with lung cancer during the trial, the median time from chest X-ray to diagnosis was 44 days with AI prioritisation and 46 days without – a two-day difference that was deemed not statistically significant. The median time from X-ray to the more detailed CT scan was identical in both groups at 53 days. The median time from chest X-ray to a fast-tracked CT scan (because cancer was suspected) was six days with AI prioritisation and seven days without, again not deemed statistically significant.
A closer examination of patients ultimately diagnosed with cancer revealed a stark pattern. When both the AI and the radiologist judged an X-ray as abnormal, the median time to cancer diagnosis was 38 days. When both judged it as normal (X-rays detect only four out of five cancers), the median time stretched to 177 days, nearly five months.
There were 53 cases where the AI flagged the X-ray as abnormal, but the radiologist’s report did not identify an abnormality. These patients waited a median of 106 days for their cancer diagnosis, significantly longer than when both agreed. The researchers say this group warrants urgent further investigation, as some could have been diagnosed at an earlier stage.
“The cases where the AI spotted something the radiologist did not flag are the ones that interest us most,” said UCLH consultant radiologist and UCL honorary associate professor Arjun Nair. “These patients waited much longer for a diagnosis. We need to understand whether there is a pattern, as this has implications not just for how we use the AI to spot these cases earlier, but also potentially how we train radiologists and radiographers of the future,” he added.
The authors argue that what is needed is not AI prioritisation alone, but a redesign of the whole pathway so that when AI flags a suspicious X-ray, a coordinated series of actions (CT booking, clinical review, specialist referral) is automatically triggered before the patient leaves the department. This will take investment in infrastructure and workforce, not just AI, the study concludes.