AI-guided lung biopsies reduce punctures and pneumothorax in a randomised trial

An AI-supported lung-biopsy system reduced the number of punctures, procedure time and pneumothorax complications in a multicentre randomised trial. GPS-Lung combines generative imaging with surgical navigation to help clinicians reach lung nodules under cone-beam CT guidance. Published on 21 October, the study measures improvements in the procedure itself rather than stopping at image quality.[1]

The system, called GPS-Lung, combines generative AI with cone-beam computed tomography, or CBCT, to support doctors positioning a biopsy needle. CBCT can produce three-dimensional images in or near the procedure room, but images may be noisy and small targets hard to locate. A needle has to approach a lesion along a viable path while avoiding vulnerable structures. The proposed navigation tool helps interpret the images and recommend a route through that constrained environment.[1]

A randomised test of the actual procedure

The registered trial, NCT06762613, compared GPS-Lung-supported biopsy with the standard CBCT-guided procedure. Average punctures declined from 2.5 to 2.0 and mean procedure time from 23.2 to 20.4 minutes. That amounts to half a puncture and 2.8 minutes less per procedure on average, calculated from the reported group means. The investigators also reported lower radiation exposure.[1]

Pneumothorax occurs when air collects around the lung. The trial reported it in 12.7% of the AI-supported group and 23.9% of the comparator group, an absolute difference of 11.2 percentage points. The reported P value was 0.033. The clinically relevant finding is a lower complication rate within this workflow; the study did not measure a cancer-survival benefit.[1]

The numerical comparison also requires attention to the study setting. Operators, lesion sizes, patient anatomy, equipment availability and local protocols influence procedural results. This trial reports performance within a defined clinical workflow. A hospital without comparable CBCT facilities, trained teams or integration would not necessarily reproduce the figures by installing software alone.[1]

The AI is planning, not performing the puncture

GPS-Lung assists image interpretation and navigation while a clinician performs the biopsy. Improving the image and proposed route can make it easier to place the needle accurately, but the operator still assesses anatomy and manages the procedure. The intervention tested is that combined workflow, so its effects cannot be assigned to the generative model alone.[1]

That combination is useful to keep in view when comparing products. A hospital would be adopting software, imaging equipment, navigation and clinical training together. Matching the trial’s facilities and operator workflow is more relevant to reproducing its results than matching a model label.[1]

The most useful follow-up is replication across operators, centres and difficult lesions, alongside the costs of integration. Fewer punctures and less radiation have direct patient value; the question for another hospital is whether it can reproduce those gains in its own procedure room.[1]

The trial makes a concrete case for image-guided AI assistance: a shorter path to the sample, with less procedural burden. Those are meaningful endpoints for a technology whose job is to help a doctor reach a small target inside a moving organ.[1]

Report an error or suggest a correction