Cooperative Learning

Please cite the following paper if you are using our model:

Qiu Z, Rivaz H, Xiao Y. Joint enhancement of automatic chest x-ray diagnosis and radiological gaze prediction with multistage cooperative learning. Med Phys. 2025 Jul;52(7):e17977. doi: 10.1002/mp.17977. PMID: 40665596; PMCID: PMC12264402.

Installation

Data

You can download the two datasets used in this study from the following links:

1. MIMIC-CXR Database
   • Contains the chest X-ray images.
   • Access requires approval through PhysioNet Credentialed Access.
2. Eye Gaze Data for Chest X-rays
   • Provides eye gaze data for chest X-ray studies.
   • Includes references to specific chest X-ray images from the MIMIC-CXR database.

Important Note: Image Matching

The eye gaze data references specific chest X-ray images from the MIMIC-CXR database. You need to manually locate and download these images from the MIMIC-CXR database one by one based on the image IDs provided in the Eye Gaze dataset.

Environment Setup

To set up the environment for this project, follow these steps:

Step 1: Clone the Repository

First, clone this repository to your local machine:

git clone https://github.com/HealthX-Lab/CXRGazeLearn.git
cd CXRGazeLearn

Step 2: Install Required Packages

Install the required packages using requirements.txt:

pip install -r requirements.txt

Dataset Processing

To process the dataset, follow these steps:

Step 1: Navigate to the Data Folder

Go to the Data folder in the project directory:

cd Data

Step 2: Update the Directory Paths

Open data.py and update the paths to your local directories for the datasets. Replace the placeholder paths with your own:

gaze_path = '/home/ziruiqiu/MscStudy/eye-gaze-dataset/physionet.org/files/egd-cxr/1.0.0'  # Replace with your own path
cxr_path = '/home/ziruiqiu/MscStudy/eye-gaze-dataset'  # Replace with your own path

Step3: Run the Data Processing Script

Execute the data.py script to process the dataset:

python data.py

Code Usage

To use the code, follow these steps. Each stage corresponds to a step in the pipeline described in the paper.

Step 1: Stage 1 - DenseNet-201 Feature Encoder

Navigate to the CL folder, which contains the code for training the DenseNet-201 feature encoder.

cd CL
python RUN.py

---

Step 2: Stage 2 - Visual Saliency Map Prediction

The second stage generates visual saliency maps that mimic the attention patterns of medical professionals.

cd ..
cd Classification
python RUN.py

---

Step 3: Stage 3 - Multi-Scale Feature-Fusion Classifier

The final stage performs the classification task using a feature-fusion approach:

cd ..
cd KL
python RUN.py

---

If you are unable to run the code, feel free to reach out leoqiuzirui@gmail.com