Add PyHealth contributions for chest X-ray analysis

[LoganCheng0525]

• Task function for classifying abnormalities (chest_xray_classification.py).
• Metrics module for radiographic agreement (radiographic_agreement.py).
• Example documentation for CheXpert dataset analysis (examples/README.md).

Chest X-Ray Analysis with PyHealth

Introduction

This example demonstrates how to use PyHealth to perform chest X-ray classification, focusing on detecting abnormalities such as pneumonia and edema. It builds on the reproducibility efforts for the UniXGen model (Lee et al., 2023), a vision-language generative model for view-specific chest X-ray generation. The example leverages the CheXpert dataset and introduces two new PyHealth contributions:

• A task function (chest_xray_classification_fn) to label chest X-rays based on diagnoses.
• A metrics module (radiographic_agreement) to evaluate inter-rater agreement for radiographic findings.

Setup

First, ensure PyHealth and its dependencies are installed. Then, import the required modules and set up logging.

import pyhealth
from pyhealth.datasets import CheXpertDataset
from pyhealth.tasks import chest_xray_classification_fn  # Import from pyhealth.tasks
from pyhealth.metrics import radiographic_agreement  # Import from pyhealth.metrics
import matplotlib.pyplot as plt
import logging
import pandas as pd
import cv2

# Set up logging
logging.basicConfig(level=logging.INFO)

# Load CheXpert dataset
dataset = CheXpertDataset(
    root="/path/to/chexpert",  # Update with actual path
    dev=False,
    refresh_cache=True
)

Notes

• Replace /path/to/chexpert with the actual path to your CheXpert dataset directory.
• Ensure the dataset is downloaded and formatted as expected by PyHealth (see CheXpert documentation).

Data Preprocessing

Use the chest_xray_classification_fn task to process the dataset and label X-ray images based on the presence of pneumonia or edema.

samples = []
for patient in dataset.patients:
    patient_samples = chest_xray_classification_fn(patient)
    samples.extend(patient_samples)

# Convert to DataFrame for analysis
df = pd.DataFrame(samples)
print(df.head())

Expected Output

The df DataFrame will contain columns like patient_id, visit_id, xray_path, view_position, and label (1 if pneumonia or edema is present, 0 otherwise).

Visualization

Visualize a sample X-ray image along with its label and view position.

sample = df.iloc[0]
img = cv2.imread(sample["xray_path"])
if img is not None:
    plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
    plt.title(f"Label: {sample['label']}, View: {sample['view_position']}")
    plt.show()
else:
    print(f"Failed to load image at {sample['xray_path']}")

Notes

• Ensure cv2 (OpenCV) is installed to load and display images.
• If the image fails to load, verify the xray_path points to a valid file.

Model Training (Simple Example)

Train a basic PyHealth model (e.g., LogisticRegression) to classify X-rays. Note that this is a placeholder; in practice, you’d likely use a deep learning model (e.g., a CNN) to extract features from X-ray images.

from pyhealth.models import LogisticRegression

model = LogisticRegression(
    feature_keys=["xray_path"],
    label_key="label",
    feature_dims=512  # Placeholder for image feature dimension
)
model.fit(dataset, batch_size=32, epochs=5)

To-Do

• Replace LogisticRegression with a more suitable model (e.g., a CNN like ResNet) and preprocess X-ray images into feature vectors.
• Adjust feature_dims based on your feature extraction method.

Evaluation

Evaluate the model’s predictions using the radiographic_agreement metric to measure inter-rater agreement between true and predicted labels.

# Placeholder for predictions (replace with actual model predictions)
y_true = [sample["label"] for sample in samples]
y_pred = model.predict(dataset)  # Adjust based on actual model

# Compute agreement
agreement_metrics = radiographic_agreement(y_true, y_pred)
print(f"Cohen's Kappa: {agreement_metrics['kappa']:.3f}")
print(f"Percent Agreement: {agreement_metrics['percent_agreement']:.2f}%")

Expected Output

• Cohen's Kappa: A value between -1 and 1, where 1 indicates perfect agreement.
• Percent Agreement: Percentage of matching labels (0-100%).

Conclusion

This example demonstrates how PyHealth can be used to classify chest X-ray abnormalities using the CheXpert dataset. The chest_xray_classification_fn task simplifies data preprocessing, while the radiographic_agreement metric provides a robust evaluation of model performance. Future work could integrate these outputs with UniXGen’s generated X-rays for enhanced analysis.

References

• Lee, Hyungyung, et al. "Vision-Language Generative Model for View-Specific Chest X-Ray Generation." arXiv preprint arXiv:2302.12172, 2023.
• PyHealth Documentation: https://pyhealth.readthedocs.io/