Abstract
AbstractObjectivesCT imaging is essential in the initial staging of lung cancer. However, free-text radiology reports do not always directly mention clinical TNM stages. We explored the capability of OpenAI’s ChatGPT to automate lung cancer staging from CT radiology reports.MethodsWe used MedTxt-RR-JA, a public de-identified dataset of 135 CT radiology reports for lung cancer. Two board-certified radiologists assigned clinical TNM stage for each radiology report by consensus. We used a part of the dataset to empirically determine the optimal prompt to guide ChatGPT. Using the remaining part of the dataset, we (i) compared the performance of two ChatGPT models (GPT-3.5 Turbo and GPT-4), (ii) compared the performance when the TNM classification rule was or was not presented in the prompt, and (iii) performed subgroup analysis regarding the T category.ResultsThe best accuracy scores were achieved by GPT-4 when it was presented with the TNM classification rule (52.2%, 78.9%, and 86.7% for the T, N, and M categories). Most ChatGPT’s errors stemmed from challenges with numerical reasoning and insufficiency in anatomical or lexical knowledge.ConclusionsChatGPT has the potential to become a valuable tool for automating lung cancer staging. It can be a good practice to use GPT-4 and incorporate the TNM classification rule into the prompt. Future improvement of ChatGPT would involve supporting numerical reasoning and complementing knowledge.Clinical relevance statementChatGPT’s performance for automating cancer staging still has room for enhancement, but further improvement would be helpful for individual patient care and secondary information usage for research purposes.Key pointsChatGPT, especially GPT-4, has the potential to automatically assign clinical TNM stage of lung cancer based on CT radiology reports.It was beneficial to present the TNM classification rule to ChatGPT to improve the performance.ChatGPT would further benefit from supporting numerical reasoning or providing anatomical knowledge.Graphical abstract
Publisher
Cold Spring Harbor Laboratory