xmindflow / ssl-contrastive

[ISBI 2024] Leveraging Unlabeled Data for 3D Medical Image Segmentation through Self-Supervised Contrastive Learning

Leveraging Unlabeled Data for 3D Medical Image Segmentation through Self-Supervised Contrastive Learning

Welcome to our GitHub repository! Our 3D semi-supervised segmentation approach addresses key challenges by leveraging two specialized subnetworks, correcting errors and enhancing contextual information. We introduce targeted verification training and self-supervised contrastive learning to improve predictions. Our model demonstrates superior performance on clinical MRI and CT scans for organ segmentation, outperforming state-of-the-art methods. Dive into our code for advanced 3D segmentation capabilities!

Please consider starring us, if you found it useful. Thanks

Updates

• November 21, 2023: First release of the code.

Quick Overview

Installation

This code has been implemented in python language using Pytorch libarary and tested in ubuntu OS, though should be compatible with related environment. following Environement and Library needed to run the code:

• CentOS Linux release 7.3.1611
• Python 3.6.13
• CUDA 9.2
• PyTorch 1.9.0
• medpy 0.4.0
• tqdm，h5py

Getting Started

Please download the prepared dataset from the following link and use the dataset path in the training and evalution code.

• LA Dataset
• Pancreas Dataset

Please change the database path and data partition file in the corresponding code.

Training

To train the network on the LA dataset, execute python pyhon train_LA. For the Pancreas dataset, use python pyhon train_pancreas

Evaluation

To evaluate the network on the LA dataset, run pyhon test_LA. For the Pancreas dataset, run pyhon test_pancreas

Citation

If you find this project useful, please consider citing:

@InProceedings{ssl-karimi,
    author    = {},
    title     = {Leveraging Unlabeled Data for 3D Medical Image Segmentation through Self-Supervised Contrastive Learning},
    booktitle = {ISBI 2024},
    month     = {},
    year      = {},
    pages     = {}
}

Acknowledgement

We build the project based on MCF-semsupervise. Thanks for their contribution.

ssl-contrastive's Issues

How should I change the code about the class_num=3 case

Sorry to bother the author. May I ask what changes I need to make to the code if I want to categorize a larger number, for example class_num=3? I would really appreciate your help. I will be grateful if you can answer!