Fewer Punctures With Generative AI vs CBCT Guidance for Lung Biopsy

A navigation system that uses generative artificial intelligence (AI) to improve cone beam CT (CBCT) imaging reduced the number of needle punctures during transthoracic image-guided lung nodule biopsy compared with conventional CBCT guidance, according to a multicenter randomized controlled trial published in npj Digital Medicine.

The investigators noted that poor CBCT image quality and the limited intelligence of surgical guidance systems have lowered the efficiency of puncture procedures, particularly for small and complex lesions. Their Generative AI-Powered Puncture Surgery Navigation System for lung punctures (GPS-Lung) generates synthetic CT images after a CBCT scan and performs surgical navigation based on those images.

In the trial (NCT06762613), conducted at three centers in China, 309 patients were enrolled, and 219 who met the inclusion criteria were randomly assigned 1:1. Of those, 110 underwent GPS-Lung–guided biopsy and 109 underwent CBCT-guided biopsy with a commercially available virtual navigation system. For the primary outcome, number of punctures, the mean was 2.0 with GPS-Lung and 2.5 with CBCT guidance (mean difference, 0.58; 95% CI, 0.18-0.97). A single puncture was needed in 39.1% and 19.3% of patients, respectively.

Mean operative time was 20.4 minutes with GPS-Lung and 23.2 minutes with CBCT guidance. Mean total radiation dose was 4357.4µGy·m² and 5570.9µGy·m², respectively (mean difference, 1213.52µGy·m²; 95% CI, 93.86-2333.18µGy·m²). Intraoperative pneumothorax occurred in 12.7% and 23.9% of patients. The groups did not differ significantly in intraoperative hemorrhage or complications within 7 days. Lesion samples were obtained from all patients in the GPS-Lung group, whereas samples from 4 patients (3.67%) in the CBCT group could not be tested.

The investigators noted that the model was validated only for lung nodule puncture and that the primary outcome was number of punctures rather than diagnostic accuracy. Preliminary diagnostic accuracy results were “non-positive,” which they said may reflect insufficient sample size. Operators could not be blinded, which may introduce residual performance bias, and the contributions of individual system components could not be separated. Four coauthors reported a financial interest in Neusoft Medical Systems.

Source: Chen L, Xiang D, Wu L, et al. A generative AI-powered navigation system for precision CBCT-guided lung biopsies: a randomized controlled trial. npj Digit Med. Published online October 21, 2026. doi:10.1038/s41746-026-03355-1