README.md

---
license: cc-by-nc-sa-4.0
language:
- en
pipeline_tag: image-segmentation
library_name: keras
tags:
- histology
- pathology
- vision
- kidney
- transplantation
- Masson's trichrome
---

# Model Card for EffNetV2S_DLV3_512_interstitial_fibrosis_kidney_transplant

## Requesting Access

Access is granted only to users whose Hugging Face primary email matches their institutional/academic email. Requests from personal domains (e.g., @gmail, @hotmail) will be declined.

## Model Description

- **Developed by:** Bertrand Chauveau, Bordeaux University Hospital
- **Model type:** Keras segmentation model, DeepLabv3+ architecture with an EfficientNetV2S backbone, channels last
- **Task:** semantic segmentation of interstitial fibrosis in kidney transplant biopsies
- **Dataset:** Trained on green Masson's trichrome, RGB tile 512 x 512 pixels at a resolution around 1 μm/pixel
- **Input:** shape (batch_size, 512, 512, 3), raw pixel values 0.-255. as float32
- **Repository:** https://github.com/bertrandchauveau/FIAT_DL
- **Paper:** pending
- **License:** cc-by-nc-sa-4.0

## Intended Use & Limitations
### Intended Use
This model is intended for automated segmentation of fibrosis in kidney transplant histopathology slides for research purposes.

### Limitations
* **Research use only**
* **Stain specificity:** The model was trained specifically on green Masson's trichrome slides from Bordeaux University Hospital


## How To Use

**Load model**
```python
import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" or "tensorflow"
import keras
from huggingface_hub import login

login()  # login with your User Access Token, found at https://huggingface.co/settings/tokens

#load model
model = keras.saving.load_model("hf://bertrandchauveau/EffNetV2S_DLV3_512_interstitial_fibrosis_kidney_transplant")
```

**Inference**
```python
from keras.utils import img_to_array, load_img
import numpy as np

img_size = (512, 512)
threshold = 0.35 #recommended threshold value

#get an example image
!wget https://raw.githubusercontent.com/bertrandchauveau/FIAT_DL/main/example_tile.jpg

#image to array
img_array = img_to_array(load_img('example_tile.jpg', target_size=img_size))
img_batch = np.expand_dims(img_array, axis=0)  #expects shape (batch, 512, 512, 3) and float32 values 0-255

#prediction
pred = model.predict(img_batch)[0]  #shape (1,512,512,2) to (512,512,2)
mask = np.where(pred[...,1] >= threshold, 255, 0).astype(np.uint8)  #pred[...,1] is the score of fibrosis, 0 of "non-fibrosis"
```

## Software Dependencies

**Python Packages**
- keras>3.0

## Contact
For any additional questions or comments, contact Bertrand Chauveau (`bertrand.chauveau@chu-bordeaux.fr`)


---
library_name: keras
---

This model has been uploaded using the Keras library and can be used with JAX,
TensorFlow, and PyTorch backends.

For more details about the model architecture, check out
[config.json](./config.json).