The TNM staging system remains the cornerstone of cancer prognosis and treatment planning (1). Accurate staging guides therapeutic strategies, predicts outcomes, and informs multidisciplinary management. In oncology practice, staging derived from PET-CT is particularly important, as this modality combines functional and structural information into a single examination (1). However, translating narrative PET-CT reports into precise TNM categories is labor-intensive and subject to variability, especially among junior clinicians. Inconsistent staging can directly influence therapeutic recommendations, clinical trial eligibility, and longitudinal outcome assessment, making reproducibility an essential requirement in daily practice (2).

Large language models (LLMs) have emerged as promising tools for medical text interpretation (2–4). The ChatGPT family in particular has attracted attention for its ability to generate coherent responses, structure radiology reports, and even pass board-style examinations (5, 6). Beyond clinician-facing tasks, LLMs have also shown strong performance in answering patient care questions in oncology settings (7). Applications in protocol selection, multilingual report translation, and automated impression generation further illustrate the versatility of these systems within radiology (3, 4). Such progress suggests that LLMs may be capable of tackling more structured and reasoning-intensive tasks, including mapping free-text PET-CT reports to standardized oncologic staging systems. Importantly, successive iterations from ChatGPT 3.5 to ChatGPT 4o and most recently ChatGPT-5 have shown marked gains in reasoning, factual consistency, and efficiency, underscoring the need to reassess performance with each new generation (8–10).

Despite these advances, the role of LLMs in systematic oncologic staging remains poorly understood. Existing studies have generally addressed individual diagnostic questions or text summarization tasks, but systematic multi-organ TNM classification has not been thoroughly examined (5, 6, 11). Moreover, comparisons across different model generations are lacking, and head-to-head benchmarking against human readers remains scarce. Whether LLMs can provide reliable staging across cancers of different organ systems, and how their performance compares with that of junior radiologists, has yet to be established.

The aim of this study was therefore to evaluate the performance of ChatGPT 5, ChatGPT 4o, and ChatGPT 3.5 in automating TNM staging from PET-CT reports across six common cancer types. Reference standards were established by senior radiologists, and results were benchmarked against junior radiologists to assess accuracy, reproducibility, and efficiency in a clinically relevant setting.

This study assessed the ability of three large language models including ChatGPT 5, ChatGPT 4o, and ChatGPT 3.5, to automate TNM staging from PET-CT reports across six common malignancies. By directly comparing different generations of models with junior radiologists, we provide evidence on the evolving role of LLMs in radiologic oncology. The results show that ChatGPT 5 achieved the highest overall accuracy, consistently surpassing ChatGPT 4o and ChatGPT 3.5, and outperforming junior radiologists in both staging accuracy and intra-observer consistency. These findings suggest that the newest generation of models may be approaching the reliability needed for clinical applications where structured interpretation of free-text reports is required.

Earlier studies of ChatGPT in radiology have mostly emphasized report structuring, protocol selection, or answering board-style questions (12–14). Few have examined its performance on systematic staging, a task that requires not only recognition of keywords but also integration of scattered observations into codified staging rules. Our results indicate that newer model generations can more reliably perform report-to-stage mapping under AJCC criteria at the case level. The contrast between ChatGPT 5 and its predecessors underscores the rapid progress in LLM capabilities, with accuracy gains evident across nearly all cancer types and staging components. In this regard, the current findings extend previous work by showing that LLMs are not limited to impression generation but can also support more logic-driven clinical tasks (15, 16).

Differences across staging components provide useful insight into where LLMs succeed and where they struggle. T staging achieved the highest accuracy, reflecting the relatively standardized language used to describe tumor size and local invasion (17). Report phrases such as “mass measuring 3.5 cm” or “extension into chest wall” are clear and map directly to staging thresholds, allowing even earlier models to perform reasonably well. In contrast, N staging posed greater challenges. Lymph node involvement is often described with variability and uncertainty, such as “borderline enlarged” or “suspicious for metastasis,” making consistent interpretation more difficult (18). This explains the lower accuracy across all models and highlights an area where automated systems may need additional refinement. M staging showed mixed results. Negative statements about distant disease were generally interpreted correctly, but scattered positive findings, particularly when presented with cautionary wording, were frequently misclassified (19). This mirrors the difficulties faced by less experienced radiologists, emphasizing that some challenges are inherent to the task itself rather than to the model alone.

Performance also varied by cancer type. Lung and breast cancers achieved the highest accuracy, likely reflecting the fact that their staging criteria are well standardized and frequently encountered in radiology reports used to train LLMs (20). Pancreatic cancer, by contrast, proved the most difficult. This is consistent with clinical experience, as pancreatic tumors often involve subtle findings such as vascular encasement or peritoneal spread that may not be consistently described in text (21). Intermediate results were seen for liver, kidney, and prostate cancers, again suggesting that the clarity and consistency of reporting language strongly influence model performance. These differences highlight the dual dependency of automated staging on both model reasoning and the structure of the source reports.

Reliability is as important as accuracy in any tool intended for clinical decision support. Intra-model agreement was strongest for ChatGPT 5, approaching the threshold of “almost perfect” agreement, while ChatGPT 4o and ChatGPT 3.5 showed progressively lower consistency (22). The stability of outputs across repeated runs is reassuring, given concerns about stochasticity in generative models (23). Interestingly, junior radiologists demonstrated lower repeatability than ChatGPT 5, underscoring that human interpretation is not free from variability even in structured tasks (24). These findings suggest that advanced LLMs may contribute not only speed but also consistency, potentially reducing inter-reader variability that has long challenged oncologic staging (25).

The clinical consequences of TNM mis-staging warrant careful consideration before deployment. Errors in N or M classification may carry disproportionate downstream impact, including inappropriate treatment intensity, incorrect eligibility assessment for clinical trials, and flawed longitudinal comparisons. These risks may be amplified by ambiguity or incomplete specification in narrative reports, which can complicate consistent mapping to AJCC staging rules. Accordingly, outputs from LLMs should be treated as decision-support suggestions rather than definitive staging. Practical guardrails can mitigate these risks by limiting use to human-verified decision support, requiring sentence-level evidence citation for each TNM component, automatically flagging uncertain or internally discordant outputs for senior review, and preserving an audit trail for quality assurance and governance.

Automated staging could provide rapid preliminary classifications, assisting junior radiologists in daily practice and accelerating preparation for multidisciplinary tumor boards (26). The substantial time savings observed with ChatGPT models compared with human readers further support their potential role in workflow optimization (27). In a practical workflow, the model could be deployed after report finalization to pre-populate structured TNM fields with sentence-level evidence from the report for rapid human verification before tumor board submission. This human-in-the-loop design preserves clinical accountability while leveraging the observed time savings and improving staging documentation consistency. More broadly, the ability to extract structured staging data from large numbers of free-text reports may facilitate clinical research, registry development, and quality monitoring (28). Importantly, these models are best conceived as augmentative rather than replacement tools. Their outputs can serve as staging suggestions that require expert confirmation, fitting naturally into the paradigm of augmented intelligence in radiology (29).

Our study has several limitations. The sample size, while spanning six malignancies, still provided limited representation of advanced stages such as T4 or N3 disease, which may restrict generalizability for these categories. The retrospective design relied on existing reports, and subtle features influencing staging could have been omitted or inconsistently described. All data were in English and derived from two institutions, which limits applicability to multilingual and more diverse clinical environments. Furthermore, we evaluated only one LLM family (ChatGPT) and did not benchmark against other competitive general-purpose or medical-oriented models. As a result, the absolute performance reported here may not directly translate to models with different training corpora, safety constraints, or inference settings, and our conclusions should be interpreted primarily as comparative evidence across ChatGPT generations. The models were tested through API calls without domain-specific fine-tuning, mirroring real-world use but leaving open the possibility of improved performance with adaptation. Finally, staging was assessed in isolation, whereas in clinical workflows decisions are integrated with pathology, laboratory, and clinical data. Future prospective evaluations in routine practice will be necessary before clinical deployment.

In conclusion, this study demonstrates that LLMs, and particularly ChatGPT 5, can achieve reliable performance in TNM staging from PET-CT reports across multiple cancer types. Accuracy was highest for T staging and lowest for nodal assessment, with clear differences across malignancies. Compared with junior radiologists, the newest model showed both higher accuracy and greater consistency, while processing cases substantially faster. Although limitations remain, these results suggest that LLMs are approaching a level of performance that could make them valuable as adjunctive tools in oncologic imaging, capable of enhancing efficiency, standardization, and reproducibility in cancer staging.