AI triage enters lung cancer imaging workflow in NHS Greater Glasgow & Clyde
Early diagnosis still shapes lung cancer outcomes, and the chest X-ray remains the first imaging test that most often opens the diagnostic pathway in the UK. In busy respiratory and primary care services, the bottleneck is no longer only image acquisition; it is the time needed for expert review when demand for reporting outstrips radiology capacity. RADICAL was designed to answer a practical question with immediate service implications: can Qure.ai’s qXR software help prioritise suspicious chest X-rays quickly enough to move patients with possible lung cancer toward follow-up CT sooner?
The study uses a stepped-wedge cluster-randomised design, combining retrospective technical evaluation with prospective clinical effectiveness work. It also extends beyond pure performance testing by incorporating qualitative assessment of acceptability and a cost-utility analysis. That mix reflects the reality of artificial intelligence in imaging: a model can only matter if it fits the reporting workflow, preserves safety, and creates value in a strained health system.
One flagged film may be enough to change the queue.
How qXR is positioned inside the reporting pathway
The intervention is built around a straightforward triage function. Standard care chest X-rays from outpatient and primary care requests are transmitted securely to qXR for interpretation, and images showing features of cancer are marked as “Urgent Suspicion of Cancer”. Those cases are then prioritised for radiologist review within the existing reporting workflow. The intended clinical gain is not to replace radiologists, but to shift the order in which images are seen when suspicious features appear on the screen.
That distinction is central to the logic of the study. Lung cancer often arrives in the system through nonspecific symptoms, and the chest X-ray can either accelerate the pathway or leave a patient waiting behind a large volume of routine studies. By surfacing the most concerning images earlier, the software is being tested as a triage layer rather than a diagnostic endpoint. The study therefore focuses on reduction in reporting time for suspicious X-rays in NHS Greater Glasgow & Clyde, while also examining whether this prioritisation is technically robust enough for clinical use.
The study period runs for 24 months, with an initial 12-month data collection phase followed by 12 months of follow-up. That extended window gives the investigators room to examine workflow behaviour over time rather than rely on a short pilot snapshot. In service-level AI studies, the durability of the process often matters as much as the raw detection task, because an algorithm that works in controlled testing can still fail when inserted into a real reporting queue.
Clinical effectiveness, safety, and workflow burden under the same lens
The primary objective is clear: assess the clinical effectiveness of qXR in prioritising patients suspected of lung cancer on chest X-ray for follow-up CT. Secondary objectives broaden the evaluation to utility, safety, technical performance, health economics, and acceptability. That broader remit is important because false reassurance, excessive flags, or poor integration into reporting pathways can erode trust quickly, even when the underlying model looks promising in retrospective analysis.
The protocol’s stepped-wedge structure is well suited to a service intervention that must be introduced across clusters over time. It allows the system to compare practice before and after implementation while still moving every cluster toward the intervention. For an imaging service, this approach mirrors the way new digital tools usually arrive in the real world: not all at once, but in stages, with operational learning accumulating as the rollout advances.
The study does not claim that AI will determine diagnosis on its own. Instead, it tests whether the software can help identify which studies deserve fastest human attention. In lung cancer pathways, even modest gains in time to review may matter if they shorten the interval between an abnormal chest X-ray and a confirmatory CT, because delays at the front end can compound through referral and staging.
Ethics approval and the regulatory frame for the trial
The protocol received REC approval in August 2023 from the North West-Greater Manchester West Research Ethics Committee, REC 23/NW/0211. The trial is registered on ClinicalTrials.gov as NCT06044454. It will follow the principles of Good Clinical Practice, and an interim report is planned for the Scottish Government. Results are expected to be presented at artificial intelligence, radiology, and respiratory meetings before publication in peer-reviewed journals.
Those governance details are not administrative afterthoughts. In a study that places algorithmic interpretation directly into a clinical workflow, ethics oversight and trial registration help define how performance, safety, and acceptability will be judged. They also create an audit trail for a technology that could affect who is reviewed first when capacity is constrained.
What the study is actually testing at the bedside
At the bedside, the relevant question is simple: does the report queue move faster for the patients who most need it? RADICAL is set up to answer that by tracing what happens after a chest X-ray is flagged as suspicious. If qXR can reliably prioritise likely cancer cases, then the pathway to CT may become more efficient without requiring a proportional increase in human reporting capacity.
There is a deeper service logic here. Lung cancer diagnosis depends on a sequence of small time-sensitive decisions, and chest X-ray interpretation sits near the front of that chain. A triage tool that can identify concerning films earlier could ease pressure on radiology teams while protecting access for patients whose imaging is most concerning. The study will also show whether clinicians and service users find the software acceptable enough to support routine use, which is a decisive issue for any AI tool intended for deployment in NHS practice.
The unanswered issue is not whether artificial intelligence can flag suspicious images in principle, but whether it can do so in a way that improves reporting flow, remains safe, and justifies its cost inside an overstretched system. RADICAL is built to test exactly that combination, and the outcome will depend on whether technical promise survives contact with everyday radiology work.