From chatbot to clinical copilot: the role of generative artificial intelligence in nuclear medicine reporting and...

Large language models (LLMs) are rapidly transforming clinical documentation workflow. This scoping review evaluates the emerging role of LLMs in nuclear medicine and molecular imaging, as recent literature suggests increasing evaluation of newer LLM architectures alongside GPT-4. We conducted a scoping review following Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews guidelines. A comprehensive search of PubMed, Lens, and Google Scholar was performed from January 2023 to January 2026. Studies were included if they applied LLMs to nuclear medicine tasks (PET/computed tomography, single-photon emission computed tomography, and scintigraphy). General radiology studies were only included if separate nuclear medicine data were extractable. A total of 14 studies were included. While the GPT-4 family remains the most studied, recent literature includes increasing evaluation of non-GPT models, with several studies reporting comparable performance in specific tasks. In data extraction tasks, such as tumor-node-metastasis staging for breast cancer, Claude 3.5 Sonnet and DeepSeek have demonstrated performance comparable or superior to GPT-4o. The applications are primarily categorized into generative reporting (drafting impressions) and data extraction (structuring unstructured text). However, limitations persist: LLMs struggle with dynamic curve analysis (e.g. renal scintigraphy) and visual interpretation. Safety remains a concern, with one study reporting an 8% rate of potentially harmful responses in patient education scenarios. LLMs are evolving from experimental chatbots into capable 'Clinical copilots' that can significantly reduce administrative burden through automated reporting and data structuring. However, because of risks of hallucinations and domain-specific limitations in dynamic imaging, human oversight remains non-negotiable. Future research must focus on multimodal integration and multicenter prospective validation.