Contrastive Pretraining for Echocardiography Segmentation with Limited Data

Published in MIUA 2022 [Paper]

This repository contains a PyTorch implementation of our method for self-supervised pretraining using SimCLR and BYOL in to improve performance on the downstream task of ventricular segmentation in ultrasound images especially with limited data.

Installation

Clone this repository and enter the directory:

git clone https://github.com/BioMedIA-MBZUAI/contrastive-echo.git
cd contrastive-echo

Make sure you have Python installed. The code is implemented for Python 3.8.10.

Install dependencies using pip:

pip install -r requirements.txt

Datasets

EchoNet-Dynamic

1. Download the dataset from EchoNet-Dynamic website
2. Run the below script to extract the images from the videos:

cd scripts/echonet
python3 extract.py

CAMUS

1. Download the dataset from CAMUS challenge website
2. Run the below script to extract the images as PNGs:

cd scripts/camus
python3 extract.py

Usage

First, create a configuration file in configs/experiment/, following the samples in the same directory.

Then you can train a model by running the experiment as follows:

python3 run.py experiment=yourconfig.yaml

Citation

@InProceedings{10.1007/978-3-031-12053-4_50,
  author="Saeed, Mohamed
  and Muhtaseb, Rand
  and Yaqub, Mohammad",
  editor="Yang, Guang
  and Aviles-Rivero, Angelica
  and Roberts, Michael
  and Sch{\"o}nlieb, Carola-Bibiane",
  title="Contrastive Pretraining for Echocardiography Segmentation with Limited Data",
  booktitle="Medical Image Understanding and Analysis",
  year="2022",
  publisher="Springer International Publishing",
  address="Cham",
  pages="680--691",
  isbn="978-3-031-12053-4"
}