A U-Net implementation of the EfficientNetV2.

EfficientV2-UNet

This package is a U-Net implementation of the EfficientNetV2, using TensorFlow.

EfficientNetV2 improves speed and parameter efficiency. This implementation also uses the ImageNet weights for training new models.

It is intended for segmentation of histological images (RGB) that are not saved in pyramidal file format (WSI).

The output segmentation are foreground / background. Multi-class segmentation is not (yet) possible.

It works on TIF images (and probably also PNG).

Installation

1. Create a python environment (e.g. with conda, python=3.9 and 3.10 work), in a CLI:

   conda create --name ev2unet_env python=3.9

2. Activate environment:

   conda activate ev2unet_env

3. GPU support for Windows (Non GPU installations not extensively tested)

   a. Install the cudatoolkit and cudnn, e.g. with conda:

   conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0

   • Windows requires a specific version of TensorFlow (i.e. v2.10.1, higher versions are not supported on Windows), which will be installed by this package.

   • Linux GPU support and Apple Silicon support will be resolved by installing this library.

4. Install this library

   pip install efficientv2-unet

5. Verify the GPU-support:

   python -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))" >> lists your active GPU

   or

   python -c "import tensorflow as tf; print(tf.test.is_gpu_available())" >> prints true if GPU enabled

   or

   ev2unet --version >> prints the versions and whether GPU support is on.

Data preparation

Mask should have background values of 0, and foreground values of 1.

At least 3 image/mask TIF pairs are required to train a model, and should be located in separate folders.

Folder Structure:

├── images
   ├── image1.tif
   ├── image2.tif
   ├── image3.tif
   └── ...
└── masks
   ├── image1.tif
   ├── image2.tif
   ├── image3.tif
   └── ...

Training a model will split the data into train, validation and test images (by default 70%, 15%, 15%, respectively). And the images will be moved to corresponding sub-folders.

Training is performed not on the full images but on tiles (with no overlap), which will be saved into corresponding sub-folders.

Usage

Command-line:

ev2unet --help

# train example:
ev2unet --train --images path/to/images --masks path/to/masks --basedir . --name myUNetName --basemodel b2 --epochs 50 --train_batch_size 32

# predict example:
ev2unet --predict --dir path/to/images --model ./models/myUnetName/myUNetName.h5 --resolution 1 --threshold 0.5

Jupyter notebooks

Examples are also available from this repository.

QuPath extension

Get the qupath-extension-efficientv2unet!

With this QuPath extension you can easily create training data and train a model via the QuPath GUI (or script). And you can also use the GUI or a script to predict.