LLM-powered TNM staging of neuroendocrine tumors from PET/CT reports

Abstract

Purpose: Imaging reports are essential for the diagnostic evaluation, treatment planning, and follow-up of patients with neuroendocrine tumors (NETs) of the gastroenteropancreatic (GEP) system. The tumor-node metastasis (TNM) classification is a common model for evaluating the prognostic value of tumor patients. However, their traditional free-text format varies in structure, detail, and clarity, leading to inconsistencies and potential omissions of critical information necessary for optimal patient management. Recent advancements in large language models (LLMs) have created new opportunities for automating complex medical assessments, including the extraction of UICC and ENETS staging classifications from imaging reports. This approach aims to improve standardization, enhance clarity, and ensure consistency, ultimately facilitating more effective multidisciplinary clinical decision-making. This study evaluates whether large language models (LLMs) can infer UICC and ENETS TNM stage for GEP‑NETs from PET/CT free‑text reports that contain descriptive findings only (no explicit TNM labels).

Methods: We evaluated several models, including ChatGPT-4o, DeepSeek V3, Claude 3.5 Sonnet, and Gemini 2.0 Flash, on a physician-generated fictitious dataset of 108 PET/CT reports with expert-annotated TNM classifications according to UICC and ENETS criteria. Model performance was assessed through F1-scores, comparing LLM-generated classifications against human expert benchmarks.

Results: Among the tested models, ChatGPT-4o demonstrated the highest accuracy, achieving microF1 scores of 0.79, 0.99 and 0.99, for T, N and M according to UICC and 0.84, 1.00 and 0.99 respectively, according to ENETS. These results indicate that LLMs have the potential to assist in oncologic staging of NETs, especially offering support for non-specialists in clinical decision-making. However, before integration into routine practice, further prospective validation and rigorous evaluation in real-world settings are necessary.

Conclusion: This study underscores the promise of LLMs in oncologic workflows while highlighting the importance of robust benchmarking and clinical validation.

Supplementary Information: The online version contains supplementary material available at 10.1186/s12880-025-02092-3.

Keywords: Clinical decision support; Large language models; Neuroendocrine tumors; PET/CT; TNM staging.

Fig. 1

(A) Schematic representation of the workflow for classifying NET UICC and ENETS TNM based on PET/CT reports. The process includes PET/CT report generation, LLM-based classification, and validation against ground truth. (B) Radar chart illustrating the hit rate for UICC TNM classification by ChatGPT-4o, DeepSeek V3, Claude 3.5 Sonnet, and Gemini 2.0 Flash. Key findings include ChatGPT-4o´s high F1 scores (0.79) for T classification. (C) shows a heatmap depicting the differences in performance for classifying T, N and M stage across models used

Fig. 2

Confusion matrices for ChatGPT-4o for the key attributes of interest. (A) UICC T stage, (B) UICC N stage and (C) UICC M stage

Fig. 3

Radar chart illustrating the hit rate for ENETS TNM classification by GPT-4o, DeepSeek V3, Claude 3.5 Sonnet, and Gemini 2.0 Flash

Fig. 4

Confusion matrices for ChatGPT-4o for the key attributes of interest. (A) ENETS T stage, (B) ENETS N stage, (C) ENETS M stage