SimCroP: Radiograph Representation Learning with Similarity-driven Cross-granularity Pre-training

Rongsheng Wang^1,2,6*, Fenghe Tang^1,2*, Qingsong Yao³, Rui Yan^1,2†, Xu Zhang^1,2,6, Zhen Huang², Haoran Lai^1,2,6, Zhiyang He⁶, Xiaodong Tao⁶, Zihang Jiang^1,2†, S. Kevin Zhou^1,2,4,5†

^*Equal contribution. ^†Corresponding author.

¹ School of Biomedical Engineering, University of Science and Technology of China 
² Suzhou Institute for Advanced Research, University of Science and Technology of China 
³ Stanford University, Palo Alto, CA, 94025, United States 
⁴ Jingsu Provincial Key Laboratory of Multimodal Digital Twin Technology, Suzhou Jiangsu, 215123, China 
⁵ Key Laboratory of Precision and Intelligent Chemistry, USTC, Hefei Anhui, 230026, China 
⁶ Anhui IFLYTEK CO., Ltd.

​

News 🥰:

• SimCroP is accepted by MICCAI 2025! 🎉

Installation

Install Python dependencies:

conda env create -f environment.yml

Data Processing

• Download CT-Rate, RadChestCT, CC-CCII, LUNA16, and BTCV dataset.
• Preprocess datasets using preprocess.py, modify the img_root and save_folder as needed.

Pre-training

We have released the code for Pre-training in our approach.

• Download the pre-trained weight of MAE and CXR-BERT
• Update the bert_checkpoint and mae_checkpoint in process_checkpoint.py with your local paths.
• Download the CSV file for pre-training here and place it under Pre-training-SimCroP/dataset.

Then, run:

cd Pre-training-SimCroP/data
python preprocess_checkpoint.py

Training

cd Pre-training-SimCroP
bash run.sh

Our pre-trained model weights are available here.

Fine-tuning

We support linear probing classification and fine-tuning segmentation for downstream tasks.

Linear Probing Classification

To evaluate classification performance with linear probing:

CUDA_VISIBLE_DEVICES=0 python train.py --name simcrop --stage train --model vit_base_patch16 --task CT-Rate --num_classes 18 \
    --pretrained_path $PATH_TO_SimCroP_CHECKPOINT \
    --dataset_path $PATH_TO_CTRATE \
    --output_dir "output/CT-Rate/SimCroP/1/" --data_volume '1' --num_steps 6000  --eval_batch_size 48 \
    --learning_rate 1.5e-3 --warmup_steps 150 --fp16 --fp16_opt_level O2 --train_batch_size 32 \
    --patience 10

• Use --task to specify the dataset for fine-tuning classification. Supported datasets: CT-Rate, RadChestCT, CC-CCII, and LUNA16.
• Use --data_volume to control the fraction of training data used.

Segmentation

To evaluate segmentation performance with fine-tuning:

CUDA_VISIBLE_DEVICES=0 python main.py --pretrained_path "../checkpoints/simcrop.pth" --output_dir "./output/LUNA16/simcrop" --out_channels 4 \
    --space_x 1.0 --space_y 1.0 --space_z 1.0 --json_path "dataset_10.json" \
    --task LUNA16 --data_dir $PATHTOLUNA16 --max_epochs 400  --val_every 20 \
    --batch_size 1 --sw_batch_size 1

• Use --task to set specific dataset for segmentation. Supported datasets: LUNA16 and BTCV.
• Use --data_volume to control the fraction of training data used.

Acknowledgement

Some codes are borrowed from ECAMP and UNETR.

Citation

If the code, paper and weights help your research, please cite:

@InProceedings{ WanRon_SimCroP_MICCAI2025,
                 author = { Wang, Rongsheng and Tang, Fenghe and Yao, Qingsong and Yan, Rui and Zhang, Xu and Huang, Zhen and Lai, Haoran and He, Zhiyang and Tao, Xiaodong and Jiang, Zihang and Zhou, S. Kevin },
                 title = { { SimCroP: Radiograph Representation Learning with Similarity-driven Cross-granularity Pre-training } }, 
                 booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
                 year = {2025},
                 publisher = {Springer Nature Switzerland},
                 volume = { LNCS 15964 },
                 month = {October},
                 pages = { 565 -- 575 },
              }

License

This project is released under the Apache 2.0 license. Please see the LICENSE file for more information.