Diffusion-based arbitrary-scale magnetic resonance image super-resolution via progressive k-space reconstruction and...

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• High-resolution Magnetic Resonance (MR) image acquisition faces hardware and time constraints.
• Existing super-resolution (SR) methods often yield over-smoothed images and are limited to fixed upsampling scales.

Purpose of the Study:

• To develop a unified diffusion-based framework for arbitrary-scale in-plane MR image super-resolution.
• To enhance MR image quality and accuracy for downstream applications.

Main Methods:

• Proposed the Progressive Reconstruction and Denoising Diffusion Model (PRDDiff) with an Adaptive Resolution Restoration Network (ARRNet).
• Implemented a multi-stage SR strategy for incremental resolution enhancement.
• Simulated downsampling using a forward diffusion process that adds Gaussian noise.

Main Results:

• PRDDiff demonstrated superior reconstruction accuracy and generalization compared to existing MR SR methods.
• The model improved downstream lesion segmentation accuracy and classification performance on clinical datasets.
• Experiments were conducted on diverse datasets including fastMRI knee/brain and clinical pediatric cerebral palsy (CP) data.

Conclusions:

• PRDDiff offers a flexible and effective solution for MR image super-resolution across arbitrary scales.
• The proposed framework enhances both image fidelity and utility for clinical analysis.
• This diffusion-based approach advances the field of medical image enhancement.