Fetal MRI Synthesis via Balanced Auto-Encoder Based Generative Adversarial Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning for image analysis necessitates extensive training data.
• Acquiring medical imaging data is costly and raises privacy concerns.
• Synthetic data can overcome limitations of real medical imaging data.

Purpose of the Study:

• To develop a Generative Adversarial Network (GAN) for synthetic fetal Magnetic Resonance Imaging (MRI) generation.
• To address the need for large, privacy-compliant datasets in fetal imaging research.
• To enable robust and efficient training of machine learning models for fetal image analysis.

Main Methods:

• Utilized an auto-encoder based Generative Adversarial Network (GAN) architecture.
• Implemented a balanced discriminator-generator power dynamic for stable training.
• Developed an approximate convergence measure for training assessment.
• Generated synthetic fetal MRI data in axial, sagittal, and coronal planes.

Main Results:

• Generated high-quality synthetic fetal MRI images.
• Demonstrated anatomical fidelity through quantitative and qualitative segmentation of fetal structures.
• Validated clinical verisimilitude by distinguishing simulated from real MRI data.
• Achieved fast and robust training of the GAN model.

Conclusions:

• The proposed GAN approach is feasible for generating realistic synthetic fetal MRI.
• Synthetic data holds significant potential for advancing machine learning in fetal imaging.
• Further research into this method is warranted for its clinical applications.