Memory-efficient image reconstruction using diffusion models for accelerated 3D non-Cartesian UTE imaging.

Purpose

Accelerated 3D non-Cartesian MRI presents unique challenges in balancing high-resolution reconstruction with computational and memory constraints. In this work, a novel, memory-efficient image reconstruction framework using score-based diffusion models tailored for highly undersampled 3D radial UTE acquisitions is proposed.

Methods

The method leverages a volumetric extension of the NCSN++ architecture with efficient NUFFT operations and integrates a conjugate gradient-based data consistency mechanism within a predictor-corrector sampling scheme.

Results

This approach enables robust and high-fidelity reconstructions from as few as 4000 spokes-representing up to 50 × undersampling relative to Nyquist. Generalizable performance across knee, ankle, and out-of-distribution shoulder datasets is demonstrated and superior image quality compared to compressed sensing and conventional iterative methods, especially at high acceleration factors is shown.

Conclusion

The proposed framework marks a significant step toward practical deployment of diffusion models in non-Cartesian MRI reconstruction.