Semi-supervised Unpaired Medical Image Segmentation Through Task-affinity Consistency

This repository contains the code implementation for the paper titled "Semi-supervised Unpaired Medical Image Segmentation Through Task-affinity Consistency." In this work, we propose a novel approach for medical image segmentation using semi-supervised learning and task-affinity consistency.

Table of Contents

• Introduction
• Getting Started
  • Prerequisites
  • Installation
• Usage
• Training
• Inference
• Citation
• Acknowledgments

Introduction

Medical image segmentation plays a critical role in various medical applications. This repository presents an approach that leverages semi-supervised learning and task-affinity consistency to improve the accuracy of medical image segmentation. The code provided here is based on the paper https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9915650.

Getting Started

Prerequisites

Before you begin, ensure you have met the following requirements:

• Python 3.6 or higher
• Pytorch version >=0.4.1.
• Other required packages (requirements.txt)

Installation

1. Clone this repository:

git clone https://github.com/jingkunchen/TAC.git

2. Install the required dependencies:

pip install -r requirements.txt

Usage

To use this code, follow these steps:

1. Download your dataset, preprocess it as needed and put the data in data/2018LA_Seg_Training Set.

2. Run the training script:

python train_class_attention.py

3. Perform inference on your test data.

Note: The release version of the code in this repository has been optimized to remove unnecessary debugging and non-essential log information. Please feel free to modify it as needed.

Training

The train_class_attention.py script serves as the main entry point for training your model. You can customize the training process by modifying the config.yaml file. This file contains all the hyperparameters and configuration settings.

Inference

To perform inference on new data, use the test script. Ensure that you have trained the model and specified the appropriate checkpoint in the configuration file.

Citation

If you find this code or our work helpful, please consider citing our paper:

@article{chen2023semi,
title={Semi-supervised unpaired medical image segmentation through task-affinity consistency},
author={Chen, Jingkun and Zhang, Jianguo and Debattista, Kurt and Han, Jungong},
journal={IEEE Transactions on Medical Imaging},
volume={42},
number={3},
pages={594--605},
year={2023},
publisher={IEEE}
}

Acknowledgments

This code is adapted from UA-MT, SASSNet, SegWithDistMap, DTC.