Related Experiment Video
Updated: Oct 25, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Bidirectional cross-modality unsupervised domain adaptation using generative adversarial networks for cardiac image
Hengfei Cui1, Chang Yuwen1, Lei Jiang1
1National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, School of Computer Science, Northwestern Polytechnical University, Xi'an, 710072, China; Centre for Multidisciplinary Convergence Computing (CMCC), School of Computer Science and Engineering, Northwestern Polytechnical University, Xi'an, 710072, China.
This study introduces a novel Generative Adversarial Networks (GAN) based framework for unsupervised domain adaptation in cardiac image segmentation. The method significantly improves segmentation accuracy and reduces errors, outperforming existing techniques.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Cardiac image segmentation faces performance degradation in unsupervised domain adaptation without ground truth labels.
- Adapting models across different imaging modalities (e.g., CT to MRI) remains a challenge.
Purpose of the Study:
- To develop a novel Generative Adversarial Networks (GAN) based bidirectional cross-modality unsupervised domain adaptation (GBCUDA) framework.
- To enhance cardiac image segmentation performance when adapting to a target domain without ground truth labels.
Main Methods:
- The GBCUDA framework utilizes GAN for image alignment and adversarial learning for feature extraction.
- A shared encoder with a self-attention mechanism and spectrum normalization stabilizes GAN training.
- Knowledge distillation loss is incorporated for processing high-level feature maps in cross-mode segmentation.
Main Results:
- The framework improved average Dice scores from 74.1% to 81.5% for cardiac substructures on CT images.
- Average symmetric surface distance (ASD) was reduced from 7.0 to 5.8 on CT images.
- For MRI images, the framework achieved an average Dice of 59.2% and reduced ASD from 5.7 to 4.9.
Conclusions:
- The proposed unsupervised domain adaptation framework demonstrates effectiveness in cross-modality cardiac image segmentation.
- The GBCUDA method shows superiority over current state-of-the-art domain adaptation techniques.
