Towards Accelerated Low-Field Multi-Contrast MRI with Anatomy-Aware Deep Unrolling

Motivation: Low-field multi-contrast MRI (MC-MRI) faces challenges with low image quality, long scanning times and a disconnect between existing imaging procedure and specific clinical needs. Goal(s): Our goal is to efficiently reconstruct high-quality MC-MR images from highly undersampled multi-contrast low-field MRI k-space data, tailored to meet specific clinical requirements. Approach: Our proposed task-oriented reconstruction model integrates learnable group sparsity and anatomy-aware regularization. The model's iterative algorithm is then unfolded into a deep network (A2MC-MRI), which jointly learn k-space sampling patterns to enhance clinically relevant regions. Results: Our A2MC-MRI outperforms existing methods, particularly in improving the imaging quality of clinically relevant targets under high acceleration. Impact: The A2MC-MRI offers a promising approach for fast MC-MRI reconstruction with enhanced task-specific imaging quality, particularly for clinically relevant anatomical regions. This method has the potential to support downstream clinical applications by catering to individual patient needs.