ReXVal Preprocessor

A simple preprocessing tool for the Radiology Report Expert Evaluation (ReXVal) dataset. This package helps process and combine the radiologist error annotations with the ground truth and candidate reports.

Overview

This tool processes two main CSV files from the REXVAL dataset:

• 50_samples_gt_and_candidates.csv: Contains ground truth and candidate reports
• 6_valid_raters_per_rater_error_categories.csv: Contains error annotations from radiologists

It combines these files and calculates mean error statistics per study and candidate type.

Installation

Requirements

• pandas

pip install pandas

Data Preparation

1. Visit the ReXVal dataset page on PhysioNet
2. Create a PhysioNet account and obtain credentials if you haven't already
3. Download the dataset files
4. Note the directory containing the CSV files mentioned above - this will be your input_path

Usage

Command Line Interface

python preprocess.py -i /path/to/dataset/directory -o /path/to/output/directory

or with long options:

python preprocess.py --input_path /path/to/dataset/directory --output_path /path/to/output/directory

Python API

from rexval_preprocessor import preprocess

# Process the dataset
df = preprocess(
    input_path="/path/to/dataset/directory",
    output_path="/path/to/output/directory"
)

Output

The tool generates a CSV file containing:

• Study ID and number
• Candidate type
• Ground truth and predicted reports
• Number of significant errors
• Number of insignificant errors
• Total number of errors

Example Output

Below is an example of the preprocessed DataFrame:

Citation

If you use this tool, please cite the original ReXVal dataset:

Please check the citation information at: https://physionet.org/content/rexval-dataset/1.0.0/