DDaTR: Dynamic Difference-aware Temporal Residual Network for Longitudinal Radiology Report Generation

Shanshan Song, Hui Tang, Honglong Yang, Xiaomeng Li

Hong Kong University of Science and Technology (HKUST)

🔨 Installation

Clone this repository and install the required packages:

git clone https://github.com/xmed-lab/DDaTR.git
cd DDaTR

conda create -n ddatr python=3.10
conda activate ddatr
pip install -r requirements.txt

🍹 Preparation

Data Acquisition

• Images: the images can be downloaded from MIMIC CXR and IU-XRay

• Annotation:

  • MIMIC CXR. Put at ./data/mimic_cxr
  • Longitudinal-MIMIC. Put at ./data/mimic_cxr
  • ReXrank test
  • IU-XRay: please download from PromptMRG

Pre-trained Weight Downloading

• Checkpoint:
  • chexbert

🍻 Quick Start for Training & Evaluation

After all the above preparation steps, you can train DDaTR with the following command:

# For MIMIC-CXR and Longitudinal-MIMIC
bash train_mimic_cxr.sh

You can directly download our trained model from DDaTR

For testing, you can use following command:

# For MIMIC-CXR and Longitudinal-MIMIC
bash test_mimic_cxr.sh

# For ReXrank
bash test_rexrank.sh

# For IU-XRay
bash test_iu_xray.sh

💙 Acknowledgement

DDaTR is built upon the awesome PromptMRG, LAVT, LDCNet.

📄 Citation

Paper link: DDaTR: Dynamic Difference-aware Temporal Residual Network for Longitudinal Radiology Report Generation.

If you use this work in your research, please cite:

@ARTICLE{11087655,
  author={Song, Shanshan and Tang, Hui and Yang, Honglong and Li, Xiaomeng},
  journal={IEEE Transactions on Medical Imaging}, 
  title={DDaTR: Dynamic Difference-Aware Temporal Residual Network for Longitudinal Radiology Report Generation}, 
  year={2025},
  volume={44},
  number={12},
  pages={5345-5357},
  keywords={Radiology report generation;Longitudinal radiology report generation;Dynamic difference-awareness;Longitudinal multimodal encoder},
  doi={10.1109/TMI.2025.3591364}}

📧 Contact

For questions and issues, please use the GitHub issue tracker or contact [ssongan@connect.ust.hk].