An Improved Contrastive Learning Network for Semi-supervised Multi-structure Segmentation in Echocardiography
This repository provides the method described in the paper
Zhaowen Qiu, et al. "An Improved Contrastive Learning Network for Semi-supervised Multi-structure Segmentation in Echocardiography"
The repository is tested on Ubuntu 20.04.6 LTS, Python 3.8, PyTorch 1.13.0 and CUDA 12.0
pip install -r requirements.txt
The repository made two improvements to the paper Semi-supervised Semantic Segmentation with Directional Context-aware Consistency, building upon the work by replacing DeeplabV3+ with U-net and modifying the structure of the projector. These changes aimed to tackle challenges in echocardiography, such as low contrast, unclear boundaries, and incomplete cardiac structures.
This repository highly depends on the CAC repository at https://github.com/dvlab-research/Context-Aware-Consistency. We thank the authors of CAC for their great work and code.
Besides, we also borrow some codes from the following repositories.
-
Firstly, download the CAMUS dataset Dataset.
-
Set the dir of your training set to 'data_dir' in the config file 'heart_cac_deeplabv3+_resnet50_1over8_datalist0.json'.
-
Set the dir of your validation set to 'val_loader' -> 'data_dir' in the config file 'heart_cac_deeplabv3+_resnet50_1over8_datalist0.json'.