(Nature 2025)

Jinxi Xiang‡, Xiyue Wang‡, Xiaoming Zhang, Yinghua Xi, Feyisope Eweje, Yijiang Chen, Yuchen Li, Colin Bergstrom, Matthew Gopaulchan, Ted Kim, Kun-Hsing Yu, Sierra Willens, Francesca Maria Olguin, Jeffrey J. Nirschl, Joel Neal, Maximilian Diehn, Sen Yang⁺, Ruijiang Li⁺ (‡Equal Contribution)

Lead Contact: Ruijiang Li, Ph.D.

Stanford University, Harvard University

---

We develop Multimodal transformer with Unified maSK modeling (MUSK), a vision-language foundation model designed to leverage large-scale, unlabeled, unpaired image-text data. MUSK is pre-trained on 50 million pathology images and 1 billion pathology-related text tokens using unified masked modeling. MUSK achieves superior performance across 23 patch-level and slide-level benchmarks, including cross-modal retrieval, visual question answering, and image classification. Importantly, MUSK shows promising performance in outcome prediction, including melanoma relapse prediction, pan-cancer prognosis prediction, and immunotherapy response prediction in lung and gastro-esophageal cancers. MUSK effectively combines complementary information from pathology images and clinical reports and can potentially improve diagnosis and precision cancer therapy.

📢 News

• Jan 8, 2025. The MUSK paper has been published in Nature.

• Nov 8, 2024. Our manuscript, "A Vision-Language Foundation Model for Precision Oncology," has been officially accepted by Nature.

• Nov 3, 2024. The initial release of the MUSK model code. The model weights will be released once the paper is published online.

Installation

First clone the repo and cd into the directory:

git clone https://github.com/lilab-stanford/MUSK
cd MUSK

Create a new enviroment with anaconda.

conda create -n musk python=3.10 -y --no-default-packages
conda activate musk
pip install --upgrade pip
pip install -r requirements.txt
pip install -e .

Model Download

The MUSK models can be accessed from HuggingFace Hub.

You need to agree to the terms to access the models and login with your HuggingFace write token:

from huggingface_hub import login
login(<huggingface write token>)

Basic Usage: MUSK as a Vision-Language Encoder

Please refer to demo.ipynb for a demonstration.

1. Load the MUSK model

import torch
from musk import utils, modeling
from timm.models import create_model
model = create_model("musk_large_patch16_384")
utils.load_model_and_may_interpolate("hf_hub:xiangjx/musk", model, 'model|module', '')
model.to(device="cuda", dtype=torch.float16)
model.eval()

2. Encode images with MUSK (refer to demo.ipynb for complete implementation)

import torchvision
from PIL import Image
from timm.data.constants import IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD

transform = torchvision.transforms.Compose([
    torchvision.transforms.Resize(384, interpolation=3, antialias=True),
    torchvision.transforms.CenterCrop((384, 384)),
    torchvision.transforms.ToTensor(),
    torchvision.transforms.Normalize(mean=IMAGENET_INCEPTION_MEAN, std=IMAGENET_INCEPTION_STD)
])

img = Image.open('assets/lungaca1014.jpeg').convert("RGB")  # input image
img_tensor = transform(img).unsqueeze(0)
with torch.inference_mode():
    image_embeddings = model(
        image=img_tensor.to("cuda", dtype=torch.float16),
        with_head=False,
        out_norm=False,
        ms_aug=True,
        return_global=True  
        )[0]  # return (vision_cls, text_cls)

• with_head=True: Enable head for image-text retrieval.
• out_norm=True: Apply normalization.
• ms_aug=True: Use multiscale augmentation (for tasks, e.g., linear probe classification, MIL).
• return_global=True: Return only [CLS] token, exclude patch tokens.

3. Encode texts with MUSK (refer to demo.ipynb for complete implementation)

tokenizer = XLMRobertaTokenizer("./musk/models/tokenizer.spm")
text = ['histopathology image of lung adenocarcinoma']
txt_ids, pad = utils.xlm_tokenizer(txt, tokenizer, max_len=100)

with torch.inference_mode():
   text_embeddings = model(
      text_description=txt_ids,
      padding_mask=pad,
      with_head=False, 
      out_norm=True,
      ms_aug=False,
      return_global=True 
   )[1]  # return (vision_cls, text_cls)

Evaluation on Patch-level Benchmarks

Please refer to ./benchmarks/demo.ipynb for a demonstration.

Patch-level benchmarks include image-text retrieval, zero-shot/few-shot/linear probe image classification, image-image retrieval, and more. The evaluation code is all-in-one which adapted from the CLIP Benchmark.

The evaluated dataset includes:

• PathMMU is available at https://huggingface.co/datasets/jamessyx/PathMMU.
• BookSet and PubmedSet are available at https://warwick.ac.uk/fac/cross_fac/tia/data/arch.
• PatchCamelyon can be accessed at https://patchcamelyon.grand-challenge.org/.
• NCT-CRC-HE-100K dataset is available at https://zenodo.org/record/1214456.
• SICAPv2 can be downloaded from https://data.mendeley.com/datasets/9xxm58dvs3/1.
• Osteo dataset is available at https://www.cancerimagingarchive.net/collection/osteosarcoma-tumor-assessment/.
• RenalCell can be downloaded from https://zenodo.org/records/6528599.
• SkinCancer is accessible at https://www.isic-archive.com/.
• LC25000 dataset is available for download at https://github.com/tampapath/lung_colon_image_set.
• PanNuke can be accessed at https://warwick.ac.uk/fac/cross_fac/tia/data/pannuke.
• UniToPatho dataset is available at https://ieee-dataport.org/open-access/unitopatho.
• WSSS4LUAD can be downloaded from https://wsss4luad.grand-challenge.org/WSSS4LUAD/.
• BRACS datasets for 3 and 6 classes are available for download at https://www.bracs.icar.cnr.it/.

To begin, download the required datasets. For demonstration purposes, we provide example datasets available here. The UniToPatho dataset has been downsampled to 224×224 resolution to enable faster downloads, as the original dataset exceeds 300GB in size. (As such, performance may vary slightly compared to the results reported in the paper.) Once downloaded, unzip the dataset to a local directory, such as /root/to/downstreams_demo. Next, update the directory path to dataset_root=/root/to/downstreams_demo. The code will then automatically extract features and carry out evaluations.

The main file is clip_benchmark.cli and includes the following options:

• --pretrained_model: Specifies the model name and the path to its weights.
• --dataset: Indicates the evaluation dataset(s); multiple datasets can be specified.
• --dataset_root: The root of datasets.
• --task: Defines the evaluation task.
• --batch_size: Sets the batch size for feature extraction.
• --output: Specifies where to save the output results.

Set the models.txt file with entries in the format: (model_name, model_path). If you want to run both MUSK and CONCH for comparison, your models.txt might look like this:

musk_large_patch16_384,hf_hub:xiangjx/musk
conch,/path/to/conch.pt

Alternatively, you can remove the CONCH entry and run MUSK alone.

Some example commands:

# >>>>>>>>>>> zero-shot image-text retrieval >>>>>>>>>>> #
 python3 -m clip_benchmark.cli eval --pretrained_model models.txt \
        --dataset   "pathmmu_retrieval"  \
        --task "zeroshot_retrieval" \
        --batch_size 256 \
        --num_workers 8 \
        --seed 42 \
        --recall_k 1 10 50 \
        --dataset_root "/root/to/downstreams_demo" \
        --output "./results/benchmark_mm_retrieval.json"

# >>>>>>>>>>> few-shot image classification >>>>>>>>>>> #
for k_shot in "${shot_list[@]}"
do
  for seed in "${seed_list[@]}"
  do
      python3 -m clip_benchmark.cli eval --pretrained_model models.txt \
          --dataset  "skin" "pannuke" "unitopatho" \
          --task "linear_probe" \
          --batch_size 256 \
          --num_workers 8 \
          --fewshot_k $k_shot \
          --seed $seed \
          --dataset_root "/root/to/downstreams_demo" \
          --output "./results/benchmark_fs_${k_shot}shot_seed${seed}.json"
  done
done

# >>>>>>>>>>> zero-shot image2image retrieval >>>>>>>>>>> #
python3 -m clip_benchmark.cli eval --pretrained_model models.txt \
        --dataset   "unitopatho_retrieval" \
        --task "image_retrieval" \
        --batch_size 256 \
        --num_workers 8 \
        --seed 41 \
        --dataset_root "/root/to/downstreams_demo" \
        --output "./results/benchmark_image_retrieval.json"

and more tasks in ./benchmarks/demo.ipynb.

Acknowledgements

The project was built on many amazing open-source repositories: Quilt1M, PathAsst, torchscale, accelerate (model pretraining), deepspeed (model pretraining), pytorch-lightning (downstream finetuning), CLIP Benchmark (model evaluation), and PORPOISE (prognosis). We thank the authors and developers for their contributions.

Issues

• Please open new threads or address questions to xiangjx@stanford.edu or xiyue.wang.scu@gmail.com

License

This model and associated code are released under the CC-BY-NC-ND 4.0 license and may only be used for non-commercial, academic research purposes with proper attribution. Any commercial use, sale, or other monetization of the MUSK model and its derivatives, which include models trained on outputs from the MUSK model or datasets created from the MUSK model, is prohibited and requires prior approval.