General-Purpose AI Takes Aim at Lung Biopsy Safety in Memorial Sloan Kettering Pilot

Artificial intelligence has moved from reading medical images to planning how doctors should pierce them. Researchers at Memorial Sloan Kettering Cancer Center have demonstrated that GPT-4—the same large language model powering consumer chatbots—can analyze CT scans and chart safe needle paths for lung biopsies, marking an early proof-of-concept for applying general-purpose AI to interventional radiology planning.

In a retrospective study of 30 consecutive lung biopsy cases published in CVIR Oncology, the AI-generated trajectories were judged clinically adequate and safe by two blinded, experienced interventional radiologists in 28 cases (93.3%). More strikingly, the AI proposed routes that crossed a median of just 5mm of lung tissue compared to 23mm for the paths actually selected by human radiologists during the original procedures—a statistically significant difference (p=0.0063) that could translate to reduced trauma and fewer complications.

From Chatbot to Surgical Planner

The research team, led by Taylor Hoffman and Francois H. Cornelis alongside Stephen B. Solomon and Debkumar Sarkar, adapted GPT-4 for medical planning by feeding it axial CT images annotated with a radial grid overlay. Using a standardized five-parameter prompt, the system required a median of just two iterations to generate complete recommendations.

The 30-case cohort included technically challenging scenarios: five apical lesions, seven subpleural nodules, three para-mediastinal targets, and four basal locations in patients with a median age of 65 years. Lesion sizes ranged from 5mm to 45mm. The original human-performed procedures resulted in minimal self-limiting pneumothorax in four cases (13.3%) and mild hemoptysis in two (6.7%), with no major complications.

Two of the 30 AI-generated paths were rejected as suboptimal. One crossed 20mm of tissue—four times the length of the human-selected route—while another unnecessarily punctured lung parenchyma for a subpleural nodule that could have been accessed more directly.

The Concordance Gap

Despite the tissue-sparing advantage, only eight of 30 cases (26.7%) showed good concordance between AI and human trajectories, defined as alignment within 10 degrees. This stands in contrast to earlier purpose-built convolutional neural networks for biopsy planning, such as the 3D system developed by Too and colleagues, which achieved 82% concordance with human choices, 2.30-degree average deviation, and over 93% sensitivity and specificity for lesion detection.

The divergence raises questions about whether GPT-4's tissue-minimizing approach represents genuinely superior planning or simply different priorities. Very short needle paths, while reducing parenchymal trauma, may compromise needle stability in coaxial biopsy techniques—a clinical nuance the feasibility study did not evaluate.

From Feasibility to Clinical Reality

The Memorial Sloan Kettering work represents an inflection point where general-purpose foundation models cross from administrative tasks into procedural planning. The study's retrospective, two-dimensional design and 30-case sample size mean substantial validation work remains before clinical deployment. Prospective trials measuring actual diagnostic yield, complication rates, procedure time, and cost-effectiveness will be essential.

For interventional radiologists, the technology could standardize trajectory selection across varying experience levels and enable pre-procedural simulation. For AI developers, the findings validate a pathway for adapting general LLMs to specialized clinical tasks without disease-specific training, potentially opening opportunities in robotic biopsy integration and real-time intraoperative guidance.

Whether general-purpose models like GPT-4 will ultimately outperform purpose-built radiology AI remains an open question—one that this feasibility study frames but does not answer.

Sources

1. AI Plans Safer Needle Paths for Lung Biopsies, Feasibility Study Finds — Scienmag.com