GAN-Based Cross-Modality Brain MRI Synthesis: Paired Versus Unpaired Training and Comparison with Diffusion and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Incomplete MRI sequences are common, hindering AI analysis.
• The effectiveness of different AI models for MRI synthesis is unclear.

Purpose of the Study:

• To evaluate cross-modality MRI synthesis (T1-to-T2) using generative adversarial networks (GANs) and diffusion models.
• To compare paired/unpaired CycleGAN, conditional denoising diffusion probabilistic model (DDPM), and transformer-enhanced GAN performance.

Main Methods:

• Utilized the BraTS 2019 brain tumor dataset for T1-to-T2 MRI translation.
• Assessed paired and unpaired CycleGAN, DDPM, and transformer-enhanced GAN.
• Evaluated inter-modality correlation (r) and Structural Similarity Index Measure (SSIM).

Main Results:

• Paired CycleGAN achieved high correlation (r≈0.92-0.93) and SSIM (≈0.90-0.92) with fast inference (<50 ms/slice).
• DDPM yielded the highest fidelity (SSIM ≈0.93-0.95, r≈0.94) but required more computation.
• Unpaired CycleGAN produced interpretable results without supervision but with lower fidelity.

Conclusions:

• Paired CycleGAN provides an efficient solution for time-sensitive clinical MRI synthesis.
• Diffusion models represent a high-fidelity but computationally intensive benchmark for MRI synthesis.
• Model choice involves a trade-off between synthesis fidelity and computational efficiency.