Physics-Driven Zero-Shot MRI Reconstruction with Non-local Image Priors

2026-06-13 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Expert, quick

Summary

A novel physics-driven Zero-Shot Self-Supervised Learning (ZS-SSL) framework is proposed for accelerated Magnetic Resonance Imaging (MRI) reconstruction, aiming to overcome supervision scarcity and optimization instability inherent in single under-sampled scan learning. This robust framework integrates physical consistency with image-domain non-local priors through three core innovations. First, a Coil Sensitivity Map (CSM)-Guided Dynamic Repository stabilizes training by filtering physically inconsistent artifacts using coil sensitivity constraints. Second, SPIRiT-based regularization enforces k-space self-consistency via a learned correlation kernel and stochastic masking. Third, a Non-Local Self-Similarity (NSS) Pixel Bank explicitly mines non-local anatomical similarities, augmenting image-domain supervision. Extensive experiments on the FastMRI dataset demonstrate that our approach achieves leading performance, particularly under high acceleration factors, effectively matching supervised methods.

Key takeaway

For Machine Learning Engineers developing accelerated MRI reconstruction models, this physics-driven ZS-SSL framework offers a robust alternative to data-hungry supervised methods. If you are struggling with overfitting or artifacts in zero-shot learning, consider integrating CSM-guided dynamic repositories and SPIRiT-based regularization. This approach allows you to achieve leading performance, even with high acceleration factors, without relying on extensive fully-sampled external datasets.

Key insights

Combining physics-driven constraints with non-local image priors enhances zero-shot MRI reconstruction stability and performance.

Principles

• Physical consistency improves ZS-SSL stability.
• K-space self-consistency enhances reconstruction.
• Non-local image priors augment supervision.

Method

The framework uses a CSM-Guided Dynamic Repository, SPIRiT-based regularization with stochastic masking, and an NSS Pixel Bank to filter artifacts, enforce k-space consistency, and leverage anatomical similarities for augmented supervision.

In practice

• Apply CSM-guided filtering for artifact reduction.
• Integrate SPIRiT regularization for k-space consistency.
• Utilize NSS Pixel Bank for image-domain supervision.

Topics

• Zero-Shot Learning
• MRI Reconstruction
• Self-Supervised Learning
• K-space Reconstruction
• Non-Local Priors
• FastMRI Dataset

Code references

• Zolento/NS-SSL

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Variational Network with Wavelet-based UNET in Accelerated MRI Reconstruction from Under Sampled K-space Data — Wavelet-based U-Net integration into a Variational Network enhances accelerated MRI reconstruction by preserving multi-scale frequency details.
• Vision-Language Model-Guided Deep Unrolling Enables Personalized, Fast MRI — PASS uses a VLM to guide personalized MRI sampling and reconstruction for faster, anomaly-aware imaging.
• A physics-informed foundation model for quantitative diffusion MRI — PIGMENT uses a physics-informed generative prior to enable reliable quantitative diffusion MRI from sparse, heterogeneous data.
• Multi-Contrast MRI Motion Correction via Parameter-Informed Disentanglement and Adaptive Experts — A unified MRI motion correction framework disentangles contrast and adapts to severity for robust, generalizable artifact removal.
• Never Seen Before: Benchmarking Genuine Zero-Shot Composed Image Retrieval with Consistent Video-Sourced Datasets — ZeroSight provides a true zero-shot CIR benchmark and a training-free MLLM method, revealing inflated existing performance.
• Learning with Semantic Priors: Stabilizing Point-Supervised Infrared Small Target Detection via Hierarchical Knowledge Distillation — VFM-driven hierarchical knowledge distillation stabilizes point-supervised ISTD by injecting robust semantics into lightweight CNNs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.