Automatic Personalized Impression Generation for PET Reports Using Large Language Models 📄✍

Authored by: Xin Tie, Muheon Shin, Ali Pirasteh, Nevein Ibrahim, Zachary Huemann, Sharon M. Castellino, Kara Kelly, John Garrett, Junjie Hu, Steve Y. Cho, Tyler J. Bradshaw

Read the full paper <!-- Link to our Arxiv paper -->

📑 Model Description

This is the domain-adapted BARTScore for evaluating the quality of PET impressions.

To check our domain-adapted text-generation-based evaluation metrics:

• BARTScore+PET
• PEGASUSScore+PET
• T5+PET

🚀 Usage

Clone this GitHub repository in a local folder

git clone https://github.com/xtie97/PET-Report-Summarization.git

Go the the folder containing codes for computing BARTScore and create a new folder called "checkpoints"

cd ./PET-Report-Summarization/evaluation_metrics/metrics/BARTScore
mkdir checkpoints
mkdir checkpoints/bart-large

Download the model weights and put them in the folder "checkpoints/bart-large". Run the code for computing text-generation-based metrics

python compute_metrics_text_generation.py

📁 Additional Resources

• Codebase for evaluation metrics: GitHub