Automatic Personalized Impression Generation for PET Reports Using Large Language Models.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Natural Language Processing in Healthcare
• Nuclear Medicine Reporting

Background:

• Personalized clinical impressions in whole-body PET reports are crucial for accurate patient care.
• Current PET reporting methods can be time-consuming, necessitating efficiency improvements.
• Large Language Models (LLMs) offer potential for automating and personalizing report generation.

Approach:

• Trained twelve LLMs on 37,370 retrospective PET reports, incorporating physician-specific reporting styles.
• Evaluated LLMs using 30 metrics, benchmarking against nuclear medicine physician quality scores.
• Selected the top-performing PEGASUS model for expert physician review of generated impressions.

Key Points:

• Domain-adapted BARTScore and PEGASUSScore best correlated with physician preferences.
• The fine-tuned PEGASUS model generated personalized impressions with 89% clinical acceptability.
• Physician-rated utility of PEGASUS-generated impressions was comparable to original reports (4.08/5 vs 4.03/5).

Conclusions:

• Fine-tuned LLMs, specifically PEGASUS, can generate clinically useful and personalized impressions for whole-body PET reports.
• This technology demonstrates significant potential to expedite the PET reporting workflow.
• Personalized AI-generated impressions may enhance efficiency without compromising diagnostic utility.