Generative AI Navigation Reduced Puncture Attempts in a Lung-Biopsy Trial

A lung biopsy asks a clinician to guide a needle toward a small target while avoiding the pleura, blood vessels and other vulnerable structures. Cone-beam CT can provide live imaging inside the procedure room, but its images are often noisier than conventional CT. That makes the path harder to see—especially for small or complex nodules.

A new multicenter randomized trial in China tested whether generative AI could make that navigation step more precise. The system, called GPS-Lung, converted cone-beam CT scans into higher-quality synthetic CT images and paired them with commercial navigation software. Among 219 participants, the AI-assisted group needed an average of 2.0 punctures per procedure, compared with 2.5 in the standard-navigation group.

That is a meaningful procedural result. It is not proof that the system diagnoses cancer more accurately, improves survival or works in every hospital.

Research references: [1]

What the randomized trial tested

The investigators first trained and validated the image-generation model using retrospective data from 3,106 people who had undergone cone-beam-CT-guided transthoracic needle biopsy at eight Chinese hospitals. The clinical test was separate: 219 participants were randomized at three centers between January 1 and March 31, 2025. Of those, 110 received biopsy guidance through the integrated GPS-Lung system and 109 received conventional cone-beam CT navigation.

The prespecified primary endpoint was the number of punctures needed during the procedure. The trial’s statistical plan targeted 220 participants, used center-stratified computer-generated randomization and calculated its sample size around that procedural endpoint—not cancer diagnosis, recovery or mortality.

The average difference was 0.58 punctures per participant, with a 95% confidence interval from 0.18 to 0.97. The result met the study’s statistical threshold (P = 0.005). A single puncture was sufficient for 39.1% of participants in the AI-assisted group and 19.3% in the control group.

The practical idea is straightforward: make the intraoperative image clearer, then help the clinician plan and follow a route to the lesion. The AI did not perform the biopsy or make the treatment decision.

Research references: [1] [2]

Several secondary measures moved in the same direction

The AI-assisted procedures were also faster on average: 20.4 minutes versus 23.2 minutes, a mean difference of 2.8 minutes. Radiation dose-area product averaged 4,357.4 versus 5,570.9 microgray-square-meters. Intraoperative pneumothorax occurred in 12.7% of the AI group and 23.9% of the control group.

Those numbers are encouraging, especially because pneumothorax is a common and consequential complication of transthoracic lung biopsy. But they were secondary outcomes. The study was powered around puncture count, and the early article-in-press record does not turn a 219-person trial into a definitive safety verdict.

The paper reported no statistically significant difference in intraoperative hemorrhage or complications assessed within seven days. The study therefore supports a narrower conclusion: in these centers, the integrated system improved several measures of procedural efficiency and produced a promising pneumothorax signal.

Research references: [1]

What the trial did not establish

The primary outcome was not diagnostic accuracy. The authors describe the diagnostic findings as preliminary and non-positive, potentially because the sample was too small for that question. A sample was obtained from every participant in the AI group, while four control-group samples could not be tested, but that does not establish that the system was better at distinguishing malignant from benign disease.

The performing team could not be fully blinded to the navigation system in use. That creates room for performance effects: clinicians may behave differently when they know they are using a new system. Participants, data collectors, outcome assessors and analysts were intended to be blinded, while the doctors and technicians carrying out the procedure necessarily knew the assigned workflow.

The intervention was also a package. GPS-Lung combined generative image reconstruction with proprietary commercial navigation software. Because the trial evaluated the integrated system, it cannot tell us how much of the result came from the synthetic CT images, the navigation layer, workflow standardization or their interaction.

And the evidence is geographically and technically bounded. The trial was conducted at three Chinese centers with cone-beam-CT-equipped interventional radiology teams. Hospitals without similar imaging hardware, trained operators or integration support should not assume the same effect.

Research references: [1] [2]

Commercial and publication context

Four coauthors disclosed financial interests in Neusoft Medical Systems, whose commercial navigation software formed part of the system. The paper states that Neusoft did not support the work. The study also acknowledges support from Xiaomi and Kingsoft and lists multiple Chinese public and foundation research grants.

These relationships do not negate the randomized result. They make independent replication and transparent comparison with alternative systems more important.

The article was published online on October 21, 2026 as a peer-reviewed, citable article in press. Nature Portfolio says this early version may receive further edits before the final version of record. Crossmark listed the document as current when checked for this review.

Research references: [3]

The human-scale takeaway

The strongest claim is modest and useful. In one multicenter randomized trial, generative-AI-enhanced navigation helped interventional teams reach lung nodules with fewer puncture attempts, shorter procedures and less radiation exposure. The pneumothorax difference deserves attention, but it needs confirmation in larger trials designed to evaluate safety.

This is what credible clinical AI progress often looks like: not an autonomous machine replacing a specialist, but a bounded tool improving the visibility and execution of one difficult step. Before broad adoption, health systems still need independent trials, diagnostic-accuracy evidence, longer safety follow-up, cost and workflow studies, and validation across different scanners, hospitals and operator experience levels.