DD-CMD | Paper link | MICCAI2026

Dual-Domain Cross-Modal Decoding for Clinical Text-Guided Medical Image Segmentation (DD-CMD) has been accepted to MICCAI2026. This is the official implementation of DD-CMD.

---

---

Environment Setup

Please create a Python 3.8 environment first, then install the required dependencies from requirements.txt.

conda create -n dd-cmd python=3.8 -y
conda activate dd-cmd
pip install -r requirements.txt

You may also use another environment manager if preferred, but the repository is prepared for Python 3.8.

Dataset Preparation

This project uses the following datasets:

1. QaTa-COV19

Original source: QaTa-COV19 on Kaggle

2. MosMedData+

Original source: MosMedData+

3. Text Annotations

The text annotations for QaTa-COV19 and MosMedData+ can be obtained from the related resources provided in LViT:

https://github.com/HUANGLIZI/LViT

We thank Li et al. for making these resources available. If you use the dataset annotations or related materials, please cite their work accordingly.

Pretrained Backbones

This repository relies on pretrained vision and language backbones. You can download them from the following links:

Vision Backbone

• ConvNeXt-Tiny: facebook/convnext-tiny-224

Language Backbone

• PubMedBERT/BiomedBERT: microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext

Configuration

Model and training settings can be modified in:

config/train.yaml

The main encoder settings used for this model are:

vision_type: facebook/convnext-tiny-224
bert_type: microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext

Please ensure the dataset paths, checkpoint paths, and any runtime options in config/train.yaml are correctly configured before launching training or evaluation.

Training

We use PyTorch Lightning for training.

Run the following command to train the model:

python train.py --config ./config/train.yaml

Evaluation

Use the following commands to evaluate the trained model.

• The pretrained model checkpoints are available and will be publicly shared later. Until then, please contact the authors or create an issue if you need access for reproduction or research purposes.

MosMedData+ Dataset

python evaluate.py --config ./config/train.yaml --ckpt ./pretrained_models/MosMedplus.ckpt

QaTa-COV19 Dataset

python evaluate.py --config ./config/train.yaml --ckpt ./pretrained_models/QaTa-Covid19.ckpt

Notes

• Ensure that all dataset files and text annotations are placed in the correct locations expected by the configuration file.
• If you use custom checkpoints, replace the --ckpt path with the path to your own saved model.
• The provided configuration file can be adjusted for batch size, learning rate, dataset paths, and other experiment settings.

Acknowledgment

We thank the authors of LViT for making their code, dataset, and related resources publicly available.

Citation

If you find this repository helpful, please consider citing our paper. When appropriate, please also cite the relevant dataset and annotation sources.

@article{rahman2026ddcmd,
  title={Dual-Domain Cross-Modal Decoding for Clinical Text-Guided Medical Image Segmentation},
  author={Rahman, Md Maklachur and Hammond, Tracy},
  journal={arXiv preprint arXiv:2608.11335},
  year={2026}
}