Cross-Modality Image Translation of 3 Tesla Magnetic Resonance Imaging to 7 Tesla Using Generative Adversarial...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Advancements in magnetic resonance imaging (MRI) present challenges in cross-modality data translation, particularly between 3 Tesla (3T) and 7 Tesla (7T) systems.
• Harmonizing data across different MRI field strengths and sites is crucial for large-scale research and clinical applications.

Purpose of the Study:

• To develop a novel method using generative adversarial networks (GANs) to synthesize 7T MRI images from 3T data.
• To evaluate the performance of a 2D CycleGAN model in accurately segmenting brain tissues (CSF, GM, WM) in synthesized 7T images.

Main Methods:

• Trained a 2D CycleGAN model on a large dataset of unpaired 3T and 7T MR images.
• Evaluated the model's performance on a paired dataset and independent testing set of 22 participants each.
• Quantified segmentation accuracy using Dice coefficient and Percentual Area Differences (PAD) for CSF, GM, and WM.

Main Results:

• The CycleGAN model successfully synthesized 7T images from 3T data, achieving high accuracy in tissue segmentation.
• Median Dice coefficients were 83.62% (CSF), 81.42% (GM), and 89.75% (WM).
• Median PAD values were 6.82% (CSF), 7.63% (GM), and 4.85% (WM), demonstrating reliable image synthesis.

Conclusions:

• The proposed GAN-based approach offers a reliable and efficient method for generating synthetic 7T MRI images from 3T data.
• This technique aids in harmonizing heterogeneous MRI datasets and has the potential to enhance contrast-to-noise ratio (CNR).
• The study highlights the utility of GANs in advancing 7T MRI research while maintaining compatibility with existing 3T data.