Assessing large language model for automated diagnosis of benign and malignant lung tumors

This is the authors' abstract. We don't add key points for this paper.

BACKGROUND: Differentiating benign from malignant lung tumors remains a critical and often challenging task in histopathological diagnosis. With the advancement of artificial intelligence (AI), large language models such as ChatGPT offer novel opportunities for assisting in diagnostic workflows. METHODS: This study retrospectively collected 250 histopathological images, including 50 cases each of lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), lung neuroendocrine neoplasms, bronchiolar adenoma, and pulmonary hamartoma. ChatGPT-4o was applied to analyze hematoxylin-eosin (HE) stained images and generate diagnostic interpretations. Its performance was compared to diagnoses made by two experienced pathologists, with evaluations based on accuracy, sensitivity, specificity, F1 score, misdiagnosis rate, and diagnostic time. RESULTS: ChatGPT-4o achieved an overall accuracy of 79.6 %, sensitivity of 81.3 %, specificity of 77.0 %, and F1 score of 82.7 %, significantly lower than the performance of human pathologists (p < 0.001). While its diagnostic accuracy for benign lesions such as bronchiolar adenoma was comparable to human experts, misclassifications frequently occurred in highly differentiated malignancies and poorly visualized neuroendocrine tumors. Notably, ChatGPT-4o achieved near real-time diagnostic speed (<10 s), vastly outperforming human diagnostic time. CONCLUSION: ChatGPT-4o shows promise as a rapid, scalable AI-assisted diagnostic tool for lung tumors, particularly for benign lesions that are morphologically challenging. Although not yet a replacement for expert pathologists, it may serve as a valuable adjunct to support decision-making and improve diagnostic efficiency in clinical pathology.