Discrete residual diffusion model for high-resolution prostate MRI synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• High-resolution magnetic resonance imaging (HR MRI) is crucial for prostate cancer (PCa) diagnosis but is limited by motion artifacts due to long scan times.
• Existing super-resolution (SR) techniques for medical images often produce overly smooth results or suffer from mode collapse.

Purpose of the Study:

• To introduce a novel generative model, the discrete residual diffusion model (DR-DM), designed to overcome limitations in current medical image super-resolution.
• To improve the clarity and diagnostic accuracy of MRI scans for prostate cancer detection.

Main Methods:

• The DR-DM employs a forward diffusion process to add noise and a backward process to learn data distribution, addressing mode collapse.
• It synthesizes residual images, focusing on high-frequency details, and uses a vector quantized variational autoencoder (VQ-VAE) to reduce computational complexity.
• A transformer architecture models the relationship between low-resolution MRI and residual images, capturing long-range dependencies for enhanced fidelity.

Main Results:

• Experimental validation on challenging datasets demonstrated the DR-DM's effectiveness compared to five state-of-the-art methods.
• The DR-DM significantly improved the signal-to-noise ratio of MRI for prostate cancer detection.

Conclusions:

• The DR-DM offers a promising solution for enhancing medical image quality, particularly for prostate cancer diagnosis.
• This approach leads to greater clarity in MRI scans, ultimately improving diagnostic accuracy for patients.