Welcome! This repository provides the official implementation of our paper TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation
Yiheng Dong, Yi Lin, Shilong Huang, Xiyan Yang, Xin Yang
In this work, we propose a Temporal Decoupling with Iterative Mutual-Refinement Model (TIM), a two-stage framework that explicitly decouples spatial pathology from temporal progression and iteratively refines reports through mutual feedback. Stage I performs temporal-decoupled representation learning, separating temporal evolution patterns from disease-specific features and generating radiology reports for both prior and current studies. Stage II introduces a mutual report refinement mechanism that identifies diagnostic inconsistencies within prior reports and iteratively rectifies both prior and current reports through error-sensitive feedback.
conda create -n TIM python=3.9
conda activate TIM
pip install -r requirements.txtMimic-cxr: you can download our preprocess annotation file from here and download the images from official website
Our training strategy consists of two stages:
-
Stage I: Temporal-decoupled Representation Learning
sh scripts/stage1_train.sh
-
Stage II: Mutual Report Refinement
sh scripts/stage2_train.sh
This project is based on R2GenGPT, we thank the original authors for their excellent work.
If you find this project useful, please consider citing:
@inproceedings{dong2026tim,
title={TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation},
author={Dong, Yiheng and Lin, Yi and Huang, Shilong and Yang, Xiyan and Yang, Xin},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={6951--6961},
year={2026}
}
