On the utility of ChatGPT in conducting a literature review on deep learning for dopamine transporter SPECT with [¹²³I]ioflupane.

Abstract

Aim: This study evaluated ChatGPT (GPT-5.2) for drafting a review paper on deep learning in dopamine transporter (DAT)-SPECT with [¹²³I]ioflupane.

Methods: The review workflow consisted of 3 steps: (i) literature search, (ii) generation of structured summaries with 24 predefined fields for each publication, and (iii) drafting a review paper based on the structured summaries across all publications. A detailed prompt for ChatGPT was iteratively designed for each step with ChatGPT support. A manual literature search was independently performed by an expert in DAT-SPECT and deep learning. ChatGPT-generated structured summaries were manually fact-checked against the full publications and corrected where necessary. The review draft generated by ChatGPT was checked against the corrected summaries.

Results: When prompted to compile an exhaustive list of publications, ChatGPT cited 13 papers, whereas the manual search identified 70 relevant publications, 67 of which were included. Corrections to ChatGPT-generated structured summaries were required in 27 cases (40.3%), affecting one or two of the 24 predefined fields, while no changes were necessary in 40 publications (59.7%). Most corrections could likely have been avoided by more precise prompting. All numerical information (dataset sizes, train-test splits, performance metrics) was correct. The review draft (~950 words) generated by ChatGPT was content-wise meaningful and accurate, but contained referencing errors, including incorrect citations, missing references, and citations of non-existent publications.

Conclusions: ChatGPT is a highly effective tool for drafting review manuscripts in nuclear medicine imaging, but its limitations in literature retrieval and referencing require careful expert supervision.