UEPS: Robust and Efficient MRI Reconstruction

A research team led by Xiang Zhou has introduced UEPS (Unrolled Expanded Progressive Sparse), a novel deep unrolled model architecture designed to solve a critical bottleneck in AI-powered MRI reconstruction. Current state-of-the-art models, known as deep unrolled models (DUMs), struggle with domain shift—their performance degrades when applied to MRI data from different machines, body parts, or clinical protocols than they were trained on. The researchers identified that estimating coil sensitivity maps (CSMs), a required step for combining signals from multiple scanner coils, is the primary limitation to generalization. UEPS tackles this with three physics-informed innovations: an Unrolled Expanded design that reconstructs each coil's image independently to bypass CSM estimation; a progressive resolution approach using k-space-to-image mapping for efficient refinement; and sparse attention mechanisms tailored to MRI's 1D undersampling patterns.

These architectural choices yield simultaneous gains in robustness and computational efficiency. To rigorously test UEPS, the team constructed a large-scale zero-shot transfer benchmark comprising 10 out-of-distribution test sets. These simulated real-world clinical shifts, including changes in anatomy (e.g., brain to knee), imaging view, contrast type, scanner vendor, magnetic field strength, and coil configurations. In extensive experiments, UEPS consistently and substantially outperformed not only other DUMs but also end-to-end, diffusion-based, and untrained neural network methods across all OOD tests. The model achieves this state-of-the-art robustness while maintaining low-latency inference, making it suitable for real-time clinical deployment where speed and reliability across diverse equipment are paramount. The work is openly available on arXiv and includes public code, advancing the path toward clinically-adoptable AI for medical imaging.