Low latency and generalizable dynamic MRI via L+S alternating gradient descent and minimization

Dynamic MRI reconstruction often requires accurate, fast, and low-latency algorithms that generalize across multiple applications, sampling schemes, and rates, without problem-specific parameter tuning. Existing methods either lack generalizability or require extensive parameter adjustment.

To develop generalizable, low-latency MRI reconstruction algorithms based on low-rank (LR) and low-rank + sparse (L+S) models that are accurate, fast, and require minimal parameter tuning across diverse dynamic MRI settings.

We propose a novel Alternating Gradient Descent and Minimization (AltGDmin) approach for L+S matrix recovery from column-wise undersampled measurements, and extend it to a 3-level hierarchical L+S model for dynamic MRI. A few-shot (FS) learning modification (AGM-L+S-FS) is introduced to enable near low-latency reconstruction with per-frame processing time comparable to or faster than scan time. A simpler LR-only variant is also evaluated.

The proposed algorithms were validated on 6 prospective and 12 retrospective dynamic MRI datasets, covering cardiac, speech, and abdominal applications, with Cartesian, pseudo-radial, radial, and spiral sampling schemes. Experiments demonstrate accurate reconstruction, generalizability across datasets and sampling rates, and low-latency performance. Detailed comparisons with state-of-the-art LR, L+S, low-latency, and unsupervised deep learning methods show superior or comparable performance.

Our AGM-L+S-FS approach provide generalizable, fast, and low-latency MRI reconstruction across multiple dynamic MRI applications and sampling schemes, without the need for problem-specific parameter tuning.