Personalized Impression Generation for PET Reports Using Large Language Models.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Natural Language Processing for Radiology

Background:

• Large language models (LLMs) show potential for summarizing radiology findings into impressions.
• Automatic impression generation for whole-body PET reports is challenging and understudied.

Purpose of the Study:

• To evaluate the clinical utility of LLMs for generating impressions in PET reports.
• To fine-tune and compare multiple open-source LLMs for this task.

Main Methods:

• Twelve open-source LLMs were fine-tuned on 37,370 retrospective PET reports.
• Models incorporated physician identity to learn specific reporting styles.
• PEGASUS was selected as the top-performing LLM based on automatic metrics and physician preference.
• Clinical utility was assessed by nuclear medicine physicians on quality dimensions and overall utility.

Main Results:

• PEGASUS demonstrated strong performance, with 89% of personalized impressions deemed clinically acceptable.
• Physicians rated personalized LLM-generated impressions with a mean utility score of 4.08/5.
• These personalized impressions were comparable in utility to physician-dictated impressions (4.03/5).

Conclusions:

• Personalized impressions generated by the PEGASUS LLM are clinically useful for PET reporting.
• LLMs have the potential to significantly expedite the drafting of PET impressions.
• This technology can assist radiologists by providing a foundation for impression generation.