PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
April 17, 2026Radiology Artificial IntelligenceOpen Access

Fine-Tuned Large Language Models for Automated Radiology Impression Generation: A Multicenter Evaluation

View Full Paper
Ask AI
Bookmark
Share

Authors

MLMingyang LiYWYue WangZMZheng Miao

Discussion

Loading...

Member takes

Overview

Multicenter evaluation demonstrates improved accuracy and efficiency of MIRA in generating radiology impressions.

Key Points

  • The aim is to develop and evaluate MIRA, a large language model, for generating accurate radiology impressions.
  • Compiled a retrospective dataset of 1.87 million radiology reports from 42 hospitals.
  • Fine-tuned MIRA using a prompt-based strategy.
  • Conducted blinded comparisons by 24 radiologists on internal and external datasets.
  • Utilized parametric and nonparametric tests for data analysis.
  • MIRA outperformed GPT-4o in both similarity and F1 score.
  • 69% of MIRA-generated impressions were rated as good as reference impressions.
  • Drafting time was reduced by 0.46 minutes per report.
  • Interradiologist agreement increased significantly.

Cite This Study

Li et al. (2026) studied this question.

synapsesocial.com/papers/69e1d0165cdc762e9d8592b7https://doi.org/10.1148/ryai.250714
View Full Paper
Ask AI
Bookmark
Share