MaCo

Key Features

This repository provides the official implementation of "Enhancing Representation in Radiography-Reports Foundation Model: A Granular Alignment Algorithm Using Masked Contrastive Learning"

In this study, we present MaCo, a masked contrastive chest X-ray foundation model that tackles these challenges. MaCo explores masked contrastive learning to simultaneously achieve fine-grained image understanding and zero-shot learning for a variety of medical imaging tasks. It designs a correlation weighting mechanism to adjust the correlation between masked chest X-ray image patches and their corresponding reports, thereby enhancing the model's representation learning capabilities. To evaluate the performance of MaCo, we conducted extensive experiments using 6 well-known open-source X-ray datasets.

Links

• Paper

Dataset Links

• MIMIC-CXR v2
• CheXpert
• NIH
• COVID Rural
• SIIM
• RSNA
• MS-CXR

Model Links

• MAE
• MaCo

Get Started

Preprocess

1. To install the required environment packages, please use the following code. It is recommended to install them in a new conda virtual environment:

pip install -r requirements.txt

2. Download the models, including the MAE weights used in the pre-training phase and the well-trained MaCo model for validating downstream tasks.

3. Download all datasets according to the "Dataset Links" and split them into training, validation, and test sets based on the "DatasetsSplits".

4. Please replace all instances of "path/to/.." in the code with the specific paths to your files.

Pre-training

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_port=29501 main_pretrain.py --SR 1 --mask_ratio 0.5 --lam 0.9 --T 0.03 --warmup_epochs 2 --batch_size 512 --epochs 51 --lr 2.5e-4 --prefix maco

Fine-tuning Classification

We provide a classification example based on the NIH dataset, including training with 1%, 10%, and 100% of the annotated data, along with the corresponding evaluation process:

cd ./CLS-NIH_ChestX-ray/
CUDA_VISIBLE_DEVICES=1 python train.py --pretrained_path /path/to/model/maco/maco.pth  --data_volume '1'   --learning_rate 8.5e-3; python test.py --model maco --gpu 1;
CUDA_VISIBLE_DEVICES=1 python train.py --pretrained_path /path/to/model/maco/maco.pth  --data_volume '10'  --num_steps 30000 --learning_rate 3e-3 --warmup_steps 500; python test.py --model maco --gpu 1;
CUDA_VISIBLE_DEVICES=1 python train.py --pretrained_path /path/to/model/maco/maco.pth  --data_volume '100' --num_steps 200000 --learning_rate 3e-3 --warmup_steps 5000;python test.py --model maco --gpu 1;

To streamline the classification process, we have included additional script files. By simply entering the necessary information, you can easily generate execution code for all three datasets.

python train_finetune_terminal_all.py

Fine-tuning Segmentation

We conduct all segmentation experiments using MMSegmentaiton (version 0.25.0). We recommend that readers set up the environment in advance and familiarize themselves with the code structure of MMSegmentation.

The necessary configuration files are provided in the "SIIM_Seg" directory. After integrating these files into the MMSegmentation framework, you can start fine-tuning using ft.sh and evaluation using test.sh.

./ft.sh
./test.sh

Fine-tuning Detection

We implement the detection task using ViTDet. We recommend that readers set up the environment in advance and familiarize themselves with the code structure of ViTDet. The necessary configuration files are provided in the "detectron2" directory.

 cd ./detectron2/projects/ViTDet
../../tools/lazyconfig_train_net_rsna.py train.init_checkpoint=path/to/maco.pth

Zero-shot classification

python Zeroshot Classification.py --pretrained_path "/path/to/maco.pth" --dataset_list ["NIH", "RSNA", "SIIM"]

Zero-shot Phrase-grounding

python Phrase Grounding --ckpt_dir "/path/to/maco.pth"

🙏 Acknowledgement

Some code is borrowed from MAE, MedKlip, HI-ML, and MRM

📝 Citation

If you find this repository useful, please consider citing this paper:

@article{huang2023enhancing,
  title={Enhancing representation in radiography-reports foundation model: A granular alignment algorithm using masked contrastive learning},
  author={Weijian Huang, Cheng Li, Hong-Yu Zhou, Hao Yang, Jiarun Liu, Yong Liang, Hairong Zheng, Shaoting Zhang, Shanshan Wang},
  journal={arXiv preprint arXiv:2309.05904},
  year={2023}
}