Diff-RRG: Longitudinal Disease-wise Patch Difference as Guidance for LLM-based Radiology Report Generation (MICCAI'25)
Overview of the proposed Diff-RRG framework.
We additionally provide a more detailed version of the overview for further details and better understanding.
Please refer to assets/method_detailed.png.
1. Prepare the code and the environment
- Git clone our repository and create a new conda environment.
cd Diff-RRG
conda create -n diffrrg python=3.9 -y
conda activate diffrrg- Install the requirements.
pip install -r requirements.txt2. Prepare the training dataset
Longitudinal-MIMIC: you can download this dataset from Here and download the images from Official website
We provide the annotation file for disease progression of Longitudinal-MIMIC dataset. You can download it from Here.
Disease progression labels. The disease_progression field is generated by comparing each study with its prior. We use the CheXbert labeler to extract disease labels for both the current and the prior study, then map the change in each disease's state as follows:
| Value | Progression state | Meaning |
|---|---|---|
0 |
stable | The state is unchanged between the prior and current study |
1 |
worsening | Absent in the prior study, present in the current study |
-1 |
improving | Present in the prior study, absent in the current study |
After downloading the data and the annotation, place them in the ./data folder.
bash scripts/1-1.mimic_train.shYou can download our pretrained Delta checkpoints from Here
bash scripts/1-2.mimic_test.shThis repository is under BSD 3-Clause License.