ReXKG

ReXKG is a system that extracts structured information from processed reports to construct a comprehensive radiology knowledge graph.

Project Structure

src/
├── data/
│   └── chexpert_plus/
│   ├── ──df_chexpert_plus_onlyfindings.csv
├── ner/
│   ├── data/
│   ├── entity/
│   ├── relation/
│   ├── shared/
│   ├── run_entity.py
│   └── run_relation.py
└── kg_construct/
    └── code/
    └── result/

System Overview

The ReXKG system consists of three main components:

1. Information Extraction System
2. Node Construction
3. Edge Construction

Information Extraction System

We use the entity extraction method proposed by PURE for our information extraction system.

Installation and Dependencies

conda env create -f environment.yml

Training

1. Data Preparation: Annotate data with GPT4, split it into train and test ./src/ner/data Run python gpt4_entity_extraction.py and python gpt4_relation_extraction.py Run python structure_data.py to convert report data into the format used by PURE for training.

2. Entity Extraction: ./src/ner Run sh run_entity.sh to train the entity extraction model.

3. Relation Extraction: ./src/ner Run sh run_relation.sh to train the relation extraction model.

4. Inference: ./src/ner/data You can also download model checkpoint from Google Drive to ./result/ Convert data file in to the test format with python get_inference_data.py ./src/ner Run sh inference.sh to perform inference on the entire dataset.

5. Data Post-processing: Run python ./result/run_relation/reverse_structure_data.py to prepare the data for node construction and edge construction.

Node Construction and Edge Construction

Installation and Dependencies

./src/kg_construct/code

1. run sh auto_build_kg.sh to get kg at result

2. We give an example of obtained kg files in ./src/kg_construct/result

Citation

If you use this code for your research or project, please cite:

@article{zhang2024uncovering,
  title={Uncovering Knowledge Gaps in Radiology Report Generation Models through Knowledge Graphs},
  author={Zhang, Xiaoman and Zhou, Hong-Yu and N. Acosta, Juli´an and Rajpurkar, Pranav},
  journal={arXiv:2408.14397},
  year={2024},
}

If you have any question, please feel free to contact.