ProTrans: Progression-aware Longitudinal Pretraining for Chest X-ray Analysis

PyTorch implementation of ProTrans, a progression-aware longitudinal pretraining framework for chest X-ray analysis.

---

Data Preparation

MIMIC-CXR-JPG

We use the MIMIC-CXR-JPG dataset as the image source. Please obtain access to MIMIC-CXR-JPG through PhysioNet: https://physionet.org/content/mimic-cxr-jpg/. Save in "./dataset".

Progression Annotation

Disease progression information is obtained from Chest ImaGenome (https://physionet.org/content/chest-imagenome/1.0.0/) and further processed to construct longitudinal training samples. Save in "./dataset"

Generate Pretraining Samples

After preparing the datasets, run:

python preprocess_code/build_pretrain_samples.py

The generated samples will be saved as JSON files for pretraining. Each pretraining sample contains:

• Prior chest X-ray and radiology report
• Current chest X-ray and radiology report
• Progression discription

---

Pretraining

Launch pretraining using:

bash scripts/1-1.mimic_train.sh

---

Model Zoo

Model	Dataset	Download
ProTrans	MIMIC-CXR	Coming Soon

---

Acknowledgements

This work is based on Diff-RRG, Med-ST, STG. We thank the creators and maintainers of these resources.

---