Skip to content

Code for the paper: Self-supervised Multi-scale Consistency for Weakly Supervised Segmentation Learning, G. Valvano et al, DART 2021

License

Notifications You must be signed in to change notification settings

gvalvano/multiscale-pyag

Repository files navigation

Self-supervised Multi-scale Consistency for Weakly Supervised Segmentation Learning

Code for the paper:

Valvano G., Leo A. and Tsaftaris S. A. (DART, 2021), Self-supervised Multi-scale Consistency for Weakly Supervised Segmentation Learning.

The official project page is here.
An online version of the paper can be found here.

Citation:

@incollection{valvano2021self,
  title={Self-supervised Multi-scale Consistency for Weakly Supervised Segmentation Learning},
  author={Valvano, Gabriele and Leo, Andrea and Tsaftaris, Sotirios A},
  booktitle={Domain Adaptation and Representation Transfer, and Affordable Healthcare and AI for Resource Diverse Global Health},
  pages={14--24},
  year={2021},
  publisher={Springer}
}

mscale_pyags


Notes:

For the experiments, refer to: experiments/acdc/exp_unet_pyag.py. This file contains the main class that is used to train on the ACDC dataset. Please, refer to the class method define_model() to see how to correctly build the CNN architecture. The structure of the segmentor can be found under the folder architectures.

Once you download the ACDC dataset and the scribble annotations, you can pre-process it using the code in the file data_interface/utils_acdc/prepare_dataset.py. You can also train with custom datasets, but you must adhere to the template required by data_interface/interfaces/dataset_wrapper.py, which assumes the access to the dataset is through a tensorflow dataset iterator.

Once preprocessed the data, you can start the training/test of the model using run.sh.

Requirements

This code was implemented using TensorFlow 1.14 and the libraries detailed in requirements.txt. You can install these libraries as: pip install -r requirements.txt or using conda (see this).

We tested the code on a TITAN Xp GPU, and on a GeForce GTX 1080, using CUDA 10.2.

About

Code for the paper: Self-supervised Multi-scale Consistency for Weakly Supervised Segmentation Learning, G. Valvano et al, DART 2021

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published