Fine-Tuned LLM Pipeline Reduces External References in Chest X-Ray Reports

An automated pipeline using a fine-tuned large language model reduced external references in MIMIC-CXR chest X-ray reports, leading to fewer context-dependent statements from a downstream vision-language model. The pipeline is open source and suitable for local deployment.

• The pipeline used parameter-efficient fine-tuning on a small set of manually revised reports.
• Reduction in external references was confirmed using both keyword-based statistics and a discriminative detection model.
• The open-source pipeline can be deployed locally, preserving data privacy.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.