Awesome

PyTorch implementation of UNet++ (Nested U-Net)

This repository contains code for a image segmentation model based on UNet++: A Nested U-Net Architecture for Medical Image Segmentation implemented in PyTorch.

[NEW] Add support for multi-class segmentation dataset.

[NEW] Add support for PyTorch 1.x.

Requirements

• PyTorch 1.x or 0.41

Installation

1. Create an anaconda environment.

conda create -n=<env_name> python=3.6 anaconda
conda activate <env_name>

2. Install PyTorch.

conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

3. Install pip packages.

pip install -r requirements.txt

Training on 2018 Data Science Bowl dataset

1. Download dataset from here to inputs/ and unzip. The file structure is the following:

inputs
└── data-science-bowl-2018
    ├── stage1_train
    |   ├── 00ae65...
    │   │   ├── images
    │   │   │   └── 00ae65...
    │   │   └── masks
    │   │       └── 00ae65...            
    │   ├── ...
    |
    ...

2. Preprocess.

python preprocess_dsb2018.py

3. Train the model.

python train.py --dataset dsb2018_96 --arch NestedUNet

4. Evaluate.

python val.py --name dsb2018_96_NestedUNet_woDS

(Optional) Using LovaszHingeLoss

1. Clone LovaszSoftmax from bermanmaxim/LovaszSoftmax.

git clone https://github.com/bermanmaxim/LovaszSoftmax.git

2. Train the model with LovaszHingeLoss.

python train.py --dataset dsb2018_96 --arch NestedUNet --loss LovaszHingeLoss

Training on original dataset

Make sure to put the files as the following structure (e.g. the number of classes is 2):

inputs
└── <dataset name>
    ├── images
    |   ├── 0a7e06.jpg
    │   ├── 0aab0a.jpg
    │   ├── 0b1761.jpg
    │   ├── ...
    |
    └── masks
        ├── 0
        |   ├── 0a7e06.png
        |   ├── 0aab0a.png
        |   ├── 0b1761.png
        |   ├── ...
        |
        └── 1
            ├── 0a7e06.png
            ├── 0aab0a.png
            ├── 0b1761.png
            ├── ...

1. Train the model.

python train.py --dataset <dataset name> --arch NestedUNet --img_ext .jpg --mask_ext .png

2. Evaluate.

python val.py --name <dataset name>_NestedUNet_woDS

Results

DSB2018 (96x96)

Here is the results on DSB2018 dataset (96x96) with LovaszHingeLoss.