Few-shot Vision-Language Model Accurately Classifies Lung Adenocarcinoma Invasiveness
Overview
This multicenter retrospective study evaluated a twenty-shot GPT-4o vision-language model for ternary classification of lung adenocarcinoma invasiveness in pure ground-glass nodules (pGGNs) on CT scans. The model demonstrated superior diagnostic accuracy compared to other training strategies and improved radiologists' performance when used as an assistive tool.
Background
Pure ground-glass nodules on chest CT scans can represent a spectrum of lung adenocarcinoma invasiveness, from preinvasive lesions to invasive adenocarcinoma. Accurate preoperative differentiation is critical for guiding surgical management but remains challenging due to overlapping imaging features and limitations of biopsy. Artificial intelligence methods, including radiomics and deep learning, have shown promise but struggle with subtle features in pGGNs. Vision-language models like GPT-4o offer a novel approach by integrating image and text data to detect key invasiveness-associated features and provide diagnostic classification.
Data Highlights
Metric
Value (95% CI)
Accuracy
0.75 (0.72–0.78)
Sensitivity
0.74 (0.70–0.77)
Specificity
0.86 (0.84–0.88)
Precision
0.75 (0.71–0.78)
F1 Score
0.74 (0.71–0.77)
AUC
0.80 (0.77–0.83)
Key Findings
The twenty-shot GPT-4o model achieved the best diagnostic performance for ternary classification of pGGN invasiveness, outperforming models trained with fewer samples (P < 0.01).
GPT-4o localized pGGNs on CT scans with a mean Dice coefficient of 0.76, indicating accurate nodule boundary delineation.
The model detected ten invasiveness-associated CT features, enabling interpretable diagnosis generation.
Radiologists rated GPT-4o outputs as highly reliable, with low risk of harm or inappropriate content.
Assistance from GPT-4o improved radiologists' diagnostic accuracy for pGGN invasiveness.
The model's diagnostic accuracy was comparable or superior to traditional radiomics and deep learning methods.
Clinical Implications
The GPT-4o vision-language model offers a noninvasive, interpretable tool to assist radiologists in accurately classifying lung adenocarcinoma invasiveness in pGGNs, potentially guiding surgical decision-making. Its ability to detect subtle CT features and improve radiologist performance may reduce reliance on invasive biopsy and enhance preoperative planning.
Conclusion
The few-shot GPT-4o vision-language model demonstrates superior capability in ternary classification of lung adenocarcinoma invasiveness in pGGNs and effectively supports radiologists in clinical diagnosis, representing a promising advancement in AI-assisted thoracic imaging.
References
GPT-4o Multicenter Study 2024 -- Assessing the Invasiveness of Lung Adenocarcinoma in Chest CT Images Using a Few-shot Vision-Language Ternary Classification Approach
