GECKO: Gigapixel Vision-Concept Contrastive Pretraining in Histopathology

Official code for our work GECKO: Gigapixel Vision-Concept Contrastive Pretraining in Histopathology

Accepted in ICCV 2025 (Highlight Poster) 🎉🎉

Requirements

To install python dependencies,

conda create -n gecko python=3.9
conda activate gecko
conda install -c conda-forge openslide
pip install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https://download.pytorch.org/whl/cu121
pip install openslide-python opencv-python scikit-image matplotlib pandas multiprocess tqdm transformers tokenizers scikit-learn numpy regex ftfy h5py timm==0.9.8

Organizing Data

Detailed description for curating data for GECKO is provided in the directory data_curation.

Training

Following curating the data as explained above, we are now ready to feed the extracted Concept Prior and Deep features for pre-training dual-branch MIL using GECKO.

Example training command:

features_dir="/your/lists_feats_dir"
experiment_dir="/your/experiment_dir"
split_path="/your/split_path_dir"

python train_gecko.py --keep_ratio 0.7 --top_k 10 --cross_val_fold 0 --dataset_dict_path "$features_dir/all_dict.pickle" --features_deep_path "$features_dir/deep_features.pth" --features_path "$features_dir/concept_prior.csv" --save_path "$experiment_dir" --split_path "$split_path" 

In train_gecko.sh and train_gecko_with_gene.sh bash file, we provide example to conduct cross-validation:

chmod +x train_gecko.sh
./train_gecko.sh /your/lists_feats_dir /your/experiment_dir /your/split_path_dir


chmod +x train_gecko_with_gene.sh
./train_gecko_with_gene.sh /your/lists_feats_dir /your/experiment_dir /your/split_path_dir /your/gene_exp_path

The Gene Expression data is provided here: Data.

Inference

After training the dual-branch MIL using GECKO, use the following notebook to explore unsupervised evaluation, supervised evaluation, and interpretability analysis.

• Notebook: evaluation.ipynb

Acknowledgements

GECKO codebase builds heavily on SI-MIL, TANGLE, CONCH, ZoomMIL, DSMIL, and CLAM. We thank the authors for their contribution.

Reported research was partially supported by the National Institutes of Health (NIH) grants 1R01CA297843-01 and 3R21CA258493-02S1. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

Reference

If you find our work useful in your research or if you use parts of this code please consider citing our paper:

@article{kapse2025gecko,
  title={GECKO: Gigapixel Vision-Concept Contrastive Pretraining in Histopathology},
  author={Kapse, Saarthak and Pati, Pushpak and Yellapragada, Srikar and Das, Srijan and Gupta, Rajarsi R and Saltz, Joel and Samaras, Dimitris and Prasanna, Prateek},
  journal={arXiv preprint arXiv:2504.01009},
  year={2025}
}