RDCNet: Instance segmentation with a minimalist recurrent residual network (MICCAI-MLMI 2020)

Official tensorflow implementation of RDCNet, a simple and efficient neural architecture for the segmentation of 2D and 3D images. For details, please refer to:

RDCNet: Instance segmentation with a minimalist recurrent residual network
Raphael Ortiz, Gustavo de Medeiros, Antoine H.F.M. Peters, Prisca Liberali, Markus Rempfler
[Paper] [Video]

Installation

This code has been tested with python 3.7, Tensorflow 2.3, Cuda 10.2, cuDNN 7.6.5 on Ubuntu 18.04

• Clone this repository

git clone https://github.com/fmi-basel/RDCNet

• Setup up a conda environment (with Cuda and cuDNN compatible for Tensorflow 2.3, adapt accordingly)

conda create -n rdcnet -c anaconda python=3.7 cudatoolkit=10.1 cudnn
conda activate rdcnet

• Install this package

pip install RDCnet/

• Run the unit tests (optional)

cd RDCNet
pytest tests/

Usage

A complete example workflow is provided as a Jupyter notebook

Instantiating a RDCNet model

A RDCNet model following the tf.keras.Model API can be instantiated by:

from rdcnet.models.rdcnet import GenericRDCnetBase
model = GenericRDCnetBase(input_shape=(None,None,None,1),
                          downsampling_factor=4,
                          n_downsampling_channels=16,
                          n_output_channels=4,
                          n_groups=8,
                          dilation_rates=(1, 2, 4, 8),
                          channels_per_group=32,
                          n_steps=5,
                          dropout=0.1)

Adding segmentation heads to an existing model

Instance segmentation heads (semi-conv embeddings + semantic class) can be added to a model having at least n_classes + n_spatial_dims output channels.

from rdcnet.models.heads import add_instance_seg_heads
model_with_heads = add_instance_seg_heads(model, n_classes=2)

Embeddings postprocessing on GPU

We also provide a GPU implementation of the Hough voting scheme to convert predicted embeddings to instance labels.

from rdcnet.postprocessing.voting import embeddings_to_labels
labels = embeddings_to_labels(embeddings,
                              foreground_mask,
                              peak_min_distance=2,
                              spacing=(1,0.23,0.23))

Hyperparameters selection

These few rules of thumb can be used to select hyperparameters for a new project

• down/up sampling factor highest factor where instances can still be visually recognized
• max dilation rate object_size << 2 * dilation_rate * downsampling_factor << training_patch_size
• network capacity primarily adjust the number of groups and channels per group
• number of iterations 5-6 is usually a good balance between performances and computational cost

Citation

If you find this work useful, please consider citing:

@inproceedings{ortiz2020,
  title={RDCNet: Instance segmentation with a minimalist recurrent residual network},
  author={Ortiz, Raphael and de Medeiros, Gustavo and Peters H.F.M., Antoine and Liberali, Prisca and Rempfler, Markus},
  booktitle={International Workshop on Machine Learning in Medical Imaging},
  year={2020},
}