Distribution matching with subset-k-space embedding for multi-contrast MRI reconstruction

Diagnostics often require multi-contrast magnetic resonance images (MC-MRI) to visualize various anatomical features. Nevertheless, equipment constraints and imaging protocols render the acquired multi-contrast image vulnerable to motion artifacts due to the long acquisition time. To reduce the time required for multiple acquisitions in MC-MRI, recent research has focused on collecting partial k-space data from a single contrast to reconstruct high-quality images by leveraging the redundancy among different contrasts. Further exploiting relevant information across diverse contrasts presents a more effective solution for accurate reconstruction.

To enhance reconstruction accuracy, this work aims to develop a novel reconstruction method that integrates the advantages of subset-k-space distribution prior and high-dimensional global prior for MC-MRI reconstruction.

Specifically, the first stage involves the individual decomposition of k-space data from different guided contrasts, which are then combined with the measurements to construct a new subset-k-space. Notably, establishing this subset-k-space minimizes the distance between the distribution of the measurements and the target examples. In addition to capitalizing on the novel distribution matching strategy for improved sampling, the second stage incorporates global prior embedding to constrain the diffusion model within the high-dimensional space, using the reconstructed contrast itself as a reference. This global prior refines the initial reconstruction obtained in the first stage.

Empirical evaluations across various datasets compellingly demonstrate the excellent capability of DMSE to preserve details and achieve accurate reconstruction.

The proposed DMSE model for MC-MRI reconstruction integrates a subset-k-space distribution prior and a high-dimensional global prior to guide the reconstruction process. By leveraging supplementary information from guidance contrasts and constrained information from the under-sampled data itself, DMSE significantly reduces noise and aliasing artifacts. Comparative and ablation experiments demonstrate that this method outperforms existing approaches in both quantitative and qualitative evaluations, achieving comparable reconstruction fidelity across different sampling conditions.