Cross-modality 3D MRI synthesis via cycle-guided denoising diffusion probability model.

Area of Science:

• Medical imaging
• Deep learning
• Magnetic Resonance Imaging (MRI)

Background:

• Cross-modality MRI synthesis is crucial for addressing missing sequences in clinical practice.
• Existing methods often face challenges in achieving high fidelity and consistency in synthesized MRIs.

Purpose of the Study:

• To introduce the cycle-guided denoising diffusion probability model (CG-DDPM), a novel deep learning framework for cross-modality MRI synthesis.
• To generate high-quality MRIs of a target modality from an existing one, improving clinical workflow and diagnostic capabilities.

Main Methods:

• The CG-DDPM framework utilizes two interconnected conditional diffusion probabilistic models.
• Cycle-guided reverse latent noise regularization is employed to enhance synthesis consistency and anatomical fidelity.
• Evaluation was performed on the BraTS2020 dataset using quantitative metrics (MSSIM, PSNR, MAE) and comparison against state-of-the-art methods (IDDPM, IDDIM, MRI-cGAN).

Main Results:

• CG-DDPM demonstrated superior performance in all cross-modality synthesis tasks (T1 → T2, T2 → T1, T1 → FLAIR, FLAIR → T1).
• Achieved highest MSSIM (0.966-0.971), lowest MAE (0.011-0.013), and competitive PSNR (27.7-28.8 dB).
• Outperformed existing methods in most metrics, showing significantly lower uncertainty and inconsistency in sampling.

Conclusions:

• The CG-DDPM offers a robust, efficient, and clinically applicable solution for cross-modality MRI synthesis.
• The framework provides improved accuracy, stability, and reduced uncertainty compared to current methods.
• Potential to streamline MRI workflows, enhance diagnostics, and support precision treatment planning in medical physics and radiation oncology.