Predicting Invasiveness of Lung Adenocarcinoma from Chest CT with Few-shot Vision-Language Ternary Classification...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Differentiating preinvasive and invasive lung adenocarcinomas in pure ground-glass nodules (pGGNs) on CT scans is clinically challenging.
• Accurate preoperative assessment is crucial for guiding treatment strategies and improving patient outcomes.

Purpose of the Study:

• To evaluate the efficacy of vision-language models, specifically GPT-4o, in noninvasively predicting the invasiveness of pGGNs.
• To assess the performance of GPT-4o compared to existing methods and its impact on radiologist diagnostic accuracy.

Main Methods:

• A retrospective multicenter study included 848 patients with pathologically confirmed lung adenocarcinoma presenting as pGGNs.
• GPT-4o was utilized to localize pGGNs, identify invasiveness-associated features, and generate diagnostic predictions, compared against Molmo.
• A twenty-shot learning approach was employed for GPT-4o.

Main Results:

• The twenty-shot GPT-4o model achieved superior performance in the ternary classification of pGGN invasiveness (P < 0.01).
• Radiologist assessments indicated high reliability and usability of GPT-4o outputs, with no significant concerns regarding harm or content.
• GPT-4o assistance led to an average improvement in diagnostic accuracy for pGGN invasiveness among radiologists.

Conclusions:

• The twenty-shot GPT-4o model demonstrates significant diagnostic capability for predicting pGGN invasiveness in lung adenocarcinoma.
• AI-assisted radiology using GPT-4o can enhance diagnostic accuracy, offering a valuable tool for preoperative assessment of lung nodules.