CheXpert Plus: Augmenting a Large Chest X-ray Dataset with Text Radiology Reports, Patient Demographics and Additional Image Formats

Table of Content

• RadGraph-XL
• CheXbert
• Model Zoo

RadGraph-XL

First install radgraph-XL through the radgraph package:

pip install radgraph

There are two ways to access the radgraph-XL annotation from the CSV. First, using row index:

import json

import pandas as pd

# Load the CSV file into a DataFrame
df = pd.read_csv('df_chexpert_plus_240401.csv')
df = df[df['section_findings'].apply(lambda x: isinstance(x, str) and len(x.split()) >= 2)]

# Load radgraph-XL annotations
with open("radgraph-XL-annotations/section_findings.json") as f:
    annotations = json.load(f)

# get fifth example
index = 5
findings = df.iloc[index]["section_findings"]
annotation = annotations[index]
print(findings)
print(annotation)

Or directly using the section content:

import json
import pandas as pd
from radgraph import utils

# Load the CSV file into a DataFrame
df = pd.read_csv('df_chexpert_plus_240401.csv')

# Load radgraph-XL annotations
with open("radgraph-XL-annotations/section_findings.json") as f:
    annotations = json.load(f)

annotations_dict = {a["0"]["text"]: a for a in annotations}

# Get a random findings:
findings = df.iloc[20]["section_findings"]
preprocessed_findings = utils.radgraph_xl_preprocess_report(findings)

# Retrieve annotation
print(annotations_dict[preprocessed_findings])

CheXbert

The json files contains the mapping between CheXpert images and diseases extracted from the radiology report section.

import json

json_diseases = [json.loads(s) for s in open("chexbert_labels/findings_fixed.json").readlines()]
print(json.dumps(json_diseases[0], indent=4))
> {
    "path_to_image": "train/patient42142/study5/view1_frontal.jpg",
    "Enlarged Cardiomediastinum": null,
    "Cardiomegaly": null,
    "Lung Opacity": null,
    "Lung Lesion": null,
    "Edema": null,
    "Consolidation": null,
    "Pneumonia": null,
    "Atelectasis": null,
    "Pneumothorax": null,
    "Pleural Effusion": null,
    "Pleural Other": null,
    "Fracture": null,
    "Support Devices": null,
    "No Finding": 1.0
}

You can merge these diseases annotations with the main CSV:

import json
import pandas as pd

# Merge both DataFrames on the 'path_to_image' column
jsonl_df = pd.read_json('chexbert_labels/findings_fixed.json', lines=True)
csv_df = pd.read_csv('df_chexpert_plus_240401.csv')
merged_df = pd.merge(jsonl_df, csv_df, on='path_to_image')

# Filter DataFrame to include only rows where 'section_findings' is not null
filtered_df = merged_df[merged_df['section_findings'].notna()]

You can further create the mapping findings -> diseases:

disease_columns = [
    "Enlarged Cardiomediastinum", "Cardiomegaly", "Lung Opacity", "Lung Lesion",
    "Edema", "Consolidation", "Pneumonia", "Atelectasis", "Pneumothorax",
    "Pleural Effusion", "Pleural Other", "Fracture", "Support Devices", "No Finding"
]

# Create a dictionary where the key is 'section_findings' and the value is mapping diseases
findings_to_diseases = {
    row['section_findings']: {disease: row[disease] for disease in disease_columns}
    for index, row in filtered_df.iterrows()
}

print(json.dumps(findings_to_diseases, indent=4))
> {
    "Unchanged right internal jugular venous catheter. Stable ....":
        {
            "Enlarged Cardiomediastinum": -1.0,
            "Cardiomegaly": NaN,
            "Lung Opacity": 1.0,
            "Lung Lesion": NaN,
            "Edema": 1.0,
            "Consolidation": NaN,
            "Pneumonia": -1.0,
            "Atelectasis": -1.0,
            "Pneumothorax": NaN,
            "Pleural Effusion": NaN,
            "Pleural Other": NaN,
            "Fracture": NaN,
            "Support Devices": 1.0,
            "No Finding": NaN
        },
    ...
}

Model Zoo

Type	Datasets	Model	Link	Tutorial
RRG	MIMIC-cxr & Chexpert Plus	Swinv2/bert-decoder-2-layers	🤗	Doc
VQGAN	MIMIC-CXR & CheXpert Plus & PadChest & BIMCV & Candid-PTX	XrayVQGAN	🤗	Doc
DINOv2	MIMIC-CXR & CheXpert Plus & PadChest & BIMCV & Candid-PTX	XrayDINOv2	🤗	Doc
CLIP	MIMIC-CXR & CheXpert Plus & PadChest & BIMCV & Candid-PTX	XrayCLIP	🤗	Doc
LLaMA	-	RadLLaMA	🤗	Doc

Reference

@article{chambon2024chexpert,
  title={CheXpert Plus: Augmenting a Large Chest X-ray Dataset with Text Radiology Reports, Patient Demographics and Additional Image Formats},
  author={Chambon, Pierre and Delbrouck, Jean-Benoit and Sounack, Thomas and Huang, Shih-Cheng and Chen, Zhihong and Varma, Maya and Truong, Steven QH and Chuong, Chu The and Langlotz, Curtis P},
  journal={arXiv preprint arXiv:2405.19538},
  year={2024}
}