Deep Learning Initialized Compressed Sensing (Deli-CS) in Volumetric Spatio-Temporal Subspace Reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Spatio-temporal MRI enables rapid whole-brain multi-parametric mapping but suffers from lengthy reconstruction times, hindering clinical use.
• Deep learning (DL) can accelerate MRI reconstruction but requires substantial computational resources and large datasets, with risks of data inconsistency.
• Enforcing data consistency is crucial for DL in MRI to prevent artifacts and ensure results align with acquired data.

Approach:

• Deep Learning Initialized Compressed Sensing (Deli-CS) is proposed to accelerate iterative MRI reconstruction.
• Deli-CS utilizes a DL-generated starting point to "kick-start" the iterative process, reducing computational load.
• The framework is applied to volumetric multi-axis spiral projection MRF for whole-brain T1 and T2 mapping at 1-mm isotropic resolution within a 2-minute acquisition.

Key Points:

• Optimized traditional reconstruction reduced time from over 2 hours to under 40 minutes with significant memory savings.
• Deli-CS further decreased full reconstruction time to 20 minutes by using 50% fewer iterations.
• The method achieves comparable reconstruction quality to traditional methods while drastically cutting down processing time.

Conclusions:

• Deli-CS effectively reduces reconstruction time for volumetric spatio-temporal MRI acquisitions.
• The approach provides a 'warm start' to iterative reconstruction algorithms, enhancing efficiency.
• This DL-based method addresses the clinical bottleneck of long reconstruction times in advanced MRI techniques.