README.md

---
tags:
- feature-extraction
- image-classification
- timm
- biology
- cancer
- owkin
- histology
library_name: timm
model-index:
- name: owkin_pancancer
  results:
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: Camelyon16[Meta]
      type: image-classification
    metrics:
    - type: accuracy
      value: 94.5 ± 4.4
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-BRCA[Hist]
      type: image-classification
    metrics:
    - type: accuracy
      value: 96.2 ± 3.3
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-BRCA[HRD]
      type: image-classification
    metrics:
    - type: accuracy
      value: 79.3 ± 2.4
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-BRCA[Mol]
      type: image-classification
    metrics:
    - type: accuracy
      value: 81.7 ± 1.6
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-BRCA[OS]
      type: image-classification
    metrics:
    - type: accuracy
      value: 64.7 ± 5.7
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-CRC[MSI]
      type: image-classification
    metrics:
    - type: accuracy
      value: 91.0 ± 2.2
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-COAD[OS]
      type: image-classification
    metrics:
    - type: accuracy
      value: 63.4 ± 7.4
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-NSCLC[CType]
      type: image-classification
    metrics:
    - type: accuracy
      value: 97.7 ± 1.3
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-LUAD[OS]
      type: image-classification
    metrics:
    - type: accuracy
      value: 53.8 ± 4.5
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-LUSC[OS]
      type: image-classification
    metrics:
    - type: accuracy
      value: 62.2 ± 2.9
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-OV[HRD]
      type: image-classification
    metrics:
    - type: accuracy
      value: 74.2 ± 8.6
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-RCC[CType]
      type: image-classification
    metrics:
    - type: accuracy
      value: 99.5 ± 0.2
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-STAD[MSI]
      type: image-classification
    metrics:
    - type: accuracy
      value: 89.9 ± 3.9
      name: ROC AUC
      verified: false
  - task:
      type: image-classification
      name: Image Classification
    dataset:
      name: TCGA-PAAD[OS]
      type: image-classification
    metrics:
    - type: accuracy
      value: 59.2 ± 4.1
      name: ROC AUC
      verified: false
widget:
- src: https://github.com/owkin/HistoSSLscaling/raw/main/assets/example.tif
  example_title: pancancer tile
co2_eq_emissions:
  emissions: 14590
  source: https://www.medrxiv.org/content/10.1101/2023.07.21.23292757v2
  training_type: pre-training
  geographical_location: Jean Zay cluster, France (~40 gCO₂eq/kWh)
  hardware_used: 32 V100 32Gb GPUs, 1216 GPU hours
license: other
pipeline_tag: feature-extraction
inference: false
datasets:
- owkin/camelyon16-features
- owkin/nct-crc-he
metrics:
- roc_auc
---

# Model card for vit_base_patch16_224.owkin_pancancer

A Vision Transformer (ViT) image classification model. \
Trained by Owkin on 40 million pan-cancer histology tiles from TCGA-COAD.

A version using the transformers library is also available here: https://huggingface.co/owkin/phikon

![](https://github.com/owkin/HistoSSLscaling/blob/main/assets/main_figure.png?raw=true)

## Model Details

- **Model Type:** Feature backbone
- **Developed by**: Owkin
- **Funded by**: Owkin and IDRIS
- **Model Stats:**
  - Params: 85.8M (base)
  - Image size: 224 x 224 x 3
  - Patch size: 16 x 16 x 3
- **Pre-training:**
  - Dataset: Pancancer40M, created from [TCGA-COAD](https://portal.gdc.cancer.gov/repository?facetTab=cases&filters=%7B%22content%22%3A%5B%7B%22content%22%3A%7B%22field%22%3A%22cases.project.project_id%22%2C%22value%22%3A%5B%22TCGA-COAD%22%5D%7D%2C%22op%22%3A%22in%22%7D%2C%7B%22content%22%3A%7B%22field%22%3A%22files.experimental_strategy%22%2C%22value%22%3A%5B%22Diagnostic%20Slide%22%5D%7D%2C%22op%22%3A%22in%22%7D%5D%2C%22op%22%3A%22and%22%7D&searchTableTab=cases)
  - Framework: [iBOT](https://github.com/bytedance/ibot), self-supervised, masked image modeling, self-distillation
- **Papers:**
  - [Scaling Self-Supervised Learning for Histopathology with Masked Image Modeling](https://www.medrxiv.org/content/10.1101/2023.07.21.23292757v2)
- **Original:** https://github.com/owkin/HistoSSLscaling
- **License:** [Owkin non-commercial license](https://github.com/owkin/HistoSSLscaling/blob/main/LICENSE.txt)

## Model Usage

### Image Embeddings
```python
from urllib.request import urlopen
from PIL import Image
import timm

# get example histology image
img = Image.open(
  urlopen(
    "https://github.com/owkin/HistoSSLscaling/raw/main/assets/example.tif"
  )
)

# load model from the hub
model = timm.create_model(
  model_name="hf-hub:1aurent/vit_base_patch16_224.owkin_pancancer",
  pretrained=True,
).eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

data = transforms(img).unsqueeze(0) # input is a (batch_size, num_channels, img_size, img_size) shaped tensor
output = model(data)  # output is a (batch_size, num_features) shaped tensor
```

## Citation
```bibtex
@article {Filiot2023.07.21.23292757,
  author = {Alexandre Filiot and Ridouane Ghermi and Antoine Olivier and Paul Jacob and Lucas Fidon and Alice Mac Kain and Charlie Saillard and Jean-Baptiste Schiratti},
  title = {Scaling Self-Supervised Learning for Histopathology with Masked Image Modeling},
  elocation-id = {2023.07.21.23292757},
  year = {2023},
  doi = {10.1101/2023.07.21.23292757},
  publisher = {Cold Spring Harbor Laboratory Press},
  URL = {https://www.medrxiv.org/content/early/2023/09/14/2023.07.21.23292757},
  eprint = {https://www.medrxiv.org/content/early/2023/09/14/2023.07.21.23292757.full.pdf},
  journal = {medRxiv}
}
```