Related Experiment Video
Updated: Jul 5, 2025
14:15
Multicolor 3D Printing of Complex Intracranial Tumors in Neurosurgery
Published on: January 11, 2020
7.1K
Multi-modal brain tumor segmentation via conditional synthesis with Fourier domain adaptation
Yasmina Al Khalil1, Aymen Ayaz1, Cristian Lorenz2
1Biomedical Engineering Department, Eindhoven University of Technology, Eindhoven, The Netherlands.
Summary
This study introduces a conditional generative adversarial network (GAN) to synthesize multi-modal brain MRI images for high-grade glioma (HGG) segmentation. Fourier domain adaptation significantly improves segmentation accuracy, addressing data scarcity in medical imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuro-oncology
Background:
- Accurate brain tumor segmentation using multi-modal MRI is crucial for diagnosis and treatment.
- Data scarcity and labeling challenges hinder the development of high-quality automated segmentation models.
- Deep learning models require extensive, expertly labeled datasets for optimal performance.
Purpose of the Study:
- To investigate the use of conditional generative adversarial networks (GANs) for synthesizing multi-modal MRI data for high-grade glioma (HGG) segmentation.
- To improve the accuracy and control of synthetic image generation by conditioning GANs on segmentation masks.
- To reduce the domain shift between synthetic and real MRI data using Fourier domain adaptation (FDA).
Main Methods:
- Developed a conditional GAN to synthesize multi-modal brain MRI images.
- Conditioned the GAN on auxiliary brain tissue and tumor segmentation masks for enhanced control.
- Applied Fourier domain adaptation (FDA) to align the frequency components of synthetic and real images.
- Trained 3D segmentation networks using both solely synthetic and augmented real/synthetic data.
Main Results:
- Fourier domain adaptation (FDA) significantly improved segmentation performance and prediction confidence.
- Models trained solely on FDA-processed synthetic data showed up to a 4% Dice score improvement.
- Augmenting real data with FDA-processed synthetic data yielded up to a 5% Dice score improvement.
- Demonstrated the effectiveness of FDA in reducing the domain shift between synthetic and real medical images.
Conclusions:
- Conditional GANs with Fourier domain adaptation offer a promising solution for generating synthetic medical images.
- This approach effectively addresses data scarcity challenges in medical image segmentation, particularly for HGG.
- Considering image frequency components is vital for successful generative models in medical image synthesis.
