You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

TEAM model weights are provided for research use. Please describe your affiliation and intended use. Do not use the model for standalone clinical diagnosis or treatment decisions.

Log in or Sign Up to review the conditions and access this model content.

TEAM

TEAM is a PyTorch model release for computational pathology feature extraction and biomarker-driven downstream prediction. It contains a patch-level pathology encoder and a slide-level TEAM aggregation checkpoint for whole-slide-image-derived patch folders.

Code repository: https://github.com/juruoxcj/TEAM

Access

This repository is gated. Please request access on Hugging Face and provide your affiliation and intended research use. Access is manually reviewed by the authors.

Model Files

File Size SHA256 Description
patch_weight.pth 1157.31 MB 1776e49a47bfb79fc2fd80aa09c88c4a6e1bcc7685a56b718ce51fbc04da127f Patch-level pathology encoder checkpoint.
slide_weight.pth 19.60 MB 495758c03c73b660946ed6c0c7c4f52c9aa06577b2f25493ec4d5df61304be60 Slide-level TEAM aggregation checkpoint.

Intended Use

TEAM is intended for research use in computational pathology workflows, including:

  • extracting patch-level pathology embeddings from image tiles;
  • aggregating patch features into slide-level TEAM representations;
  • supporting downstream biomarker-driven prognostic stratification and therapeutic response prediction experiments when paired with the released source code.

TEAM is not intended for standalone clinical diagnosis, treatment selection, or deployment as a medical device.

Quick Start

Install the code from GitHub:

git clone https://github.com/juruoxcj/TEAM.git
cd TEAM
conda env create -f environment.yml
conda activate team

After your Hugging Face access request is approved, download the weights and run the demo:

python run_team.py \
  --config ./configs/demo_config.json \
  --patch_ckpt ./patch_weight.pth \
  --slide_ckpt ./slide_weight.pth \
  --device cuda

The included GitHub demo uses six unmodified TCGA BRCA patch images to demonstrate the expected input layout. On the tested CUDA workstation, the demo completed in about 21 seconds.

Inputs

TEAM expects one folder per slide, containing pre-extracted image patches:

slides/
`-- slide_A/
    |-- patch_0001.png
    |-- patch_0002.png
    `-- ...

Supported image extensions are .png, .jpg, .jpeg, .tif, .tiff, .bmp, and .webp.

Optional de-identified clinical text can be provided as a JSON file mapping slide folder names to text. Do not include protected health information in filenames, metadata, or text fields.

Outputs

The upstream inference script writes a PyTorch .pt file containing:

Key Shape Description
patch_feat [N, 1024] Patch-level embeddings for N input patches.
slide_feat [1, 512] Slide-level TEAM embedding.
paths length N Input patch paths in feature order.

Tested Environment

  • Python 3.10.18
  • PyTorch 2.7.1+cu126
  • CUDA runtime 12.6
  • NVIDIA RTX 5880 Ada Generation GPU
  • Linux 5.15 x86_64

CPU execution is suitable for installation checks and downstream smoke tests. GPU execution is recommended for upstream feature extraction.

Training Data and Evaluation

TEAM was developed for pathology image analysis using whole-slide-image-derived patch data and associated study metadata described in the accompanying manuscript. Raw WSIs, patient-derived patch tiles, clinical records, patient-level annotations, and generated feature tensors are not redistributed in this model repository.

Quantitative evaluation should be interpreted through the accompanying manuscript and its Data Availability statement. The small public demo patches in the GitHub repository are provided only to demonstrate the input layout and inference workflow; they are not an evaluation dataset.

Limitations and Ethics

  • The released checkpoints are for research use.
  • Performance can vary across scanners, staining protocols, tissue preparation pipelines, and patient cohorts.
  • Clinical deployment requires independent validation, governance review, and compliance with local regulations.
  • Users are responsible for following data-use agreements and privacy requirements for controlled pathology data.

License

The model card and associated code release use the Apache License 2.0. Checkpoint use should follow this license and any applicable data-use restrictions described by the authors.

Citation

@article{mao2026interactive,
  title={An Interactive Trustworthy AI Pathology Copilot to Improve Biomarker-Driven Prognostic Stratification and Therapeutic Response Prediction},
  author={Mao, Yixiao and Xie, Chengjie and Li, Feng and Li, Danyi and Zhang, Wenyan and Zhang, Yidan and Li, Bingbing and Zhao, Chenglong and Zhang, Zhengyu and Tan, Ying and others},
  journal={medRxiv},
  pages={2026--05},
  year={2026},
  publisher={Cold Spring Harbor Laboratory Press}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support