Identifying Confabulation Hotspots of Large Language Models in Radiology through Semantic Entropy

Introduction

This repository contains the code used for the (not yet published) paper titled "[TODO ADD PAPER]". If you find this code or research valuable for your work, please consider citing our paper: [TODO ADD PAPER]

We have created a set of 90 questions, which span 4 different areas:

• Guidelines and Indications
• Image Acquisition
• Imaging Education
• Research

The CSV files containing all the questions can be found in the questions folder.

Requirements

Note: We used Ubuntu 22.04

1. Ensure conda is installed on your system.
2. Set up the conda environment using the provided environment.yml file:

$ conda env create --name semantic_entropy_radiology --file=environment.yml

3. Activate the environment:

$ conda activate semantic_entropy_radiology

The environment setup is now complete.

Quick Evaluation

To reproduce our results, run:

$ python3 eval.py

This command processes the cached LLM responses (stored in cache) to regenerate our findings. For a complete evaluation including new LLM prompts, see the "Reproducing the Results" section below.

Reproducing the Results

Note: Due to the non-deterministic nature of LLMs, new prompting may yield slightly different results.

1. Configure credentials:

   • Open CONFIG.py
   • Update LOCAL_KEY, LOCAL_ENDPOINT, OPENAI_KEY and AZURE_ENDPOINT with your credentials.

2. Prepare the environment:

   • Clear the cache directory to ensure a clean evaluation.

3. Generate new answers:

$ python3 generateAnswers.py

Results are saved in the cache directory with filenames prefixed with EVAL.

4. Run the evaluation:

$ python3 eval.py

This step utilizes GPT4o for semantic clustering and answer comparison. Previous entailment results are cached in entailmentCacheFile_GPT4o.csv.

Project Components

Configuration Files

• CONFIG.py: Contains the configurations needed to run the code locally.

Dataset Structure

• RadDataset.py: Defines the dataset structure used in the project.

Core Functionality

Answer Generation

• generateAnswers.py: Handles answer generation from the dataset
  • Generates primary answers (ID 0) with temperature 0.1 for accuracy assessment
  • Creates additional answers for semantic entropy computation
  • Stores results in designated output files

LLM Integration

• promptLLM.py: Implements the logic for prompting various large language models (LLMs).

Semantic Entropy Calculation

• clusterAnswers.py: Implements semantic entropy calculations and evaluation metrics (AUROC, AURAC)
• EntailmentCheck.py: Manages clustering and entailment verification
  • Implements prompt caching for efficient repeated executions

Radiologist correction

To verify answer correctness, all correctness labels were reviewed by a board-certified radiologist. This led to changes in some reported values. The corresponding corrections are included in entailmentCacheFile_GPT4o.csv.

Evaluation Tools

• eval.py: Handles the evaluation pipeline
• Features bootstrapping analysis for confidence interval calculations across semantic entropy thresholds