Automatic Impression Generation for Positron Emission Tomography Reports using Lightweight Adaptation of Pretrained Large Language Models 📑

This repository contains the code for our team project (member: Nuohao Liu, Xin Tie, Xiaogeng Liu) as part of the CS776 Advanced Natural Language Processing course. Check our presentation 📜.

Overview 🔎

Background: Adapting LLMs for PET report summarization can be quite expensive in terms of computational time and memory useage. Parameter Efficient Fine-Tuning (PEFT) presents a promising alternative that could retain high performance while utilizing fewer training resources. In this project, we aim to evaluate the effectiveness of PEFT in fine-tuning LLMs for summarizing PET findings. Our ultimate goal is to address the challenge of increasing memory when training a LLM for summarizing multiple radiology reports.

Usage 🚀

We investigated three PEFT techniques:

• LoRA
• (IA)3
• Prompt tuning

The training was powered by deepspeed

To run the training

python finetune_FlanT5.py

To run the prediction

python predict_FlanT5.py

To test the output impressions using ROUGE

python compute_rouge.py