A U-Net Transformer for magnetic resonance elastography.
Authors
Affiliations (5)
Affiliations (5)
- Department of Civil and Environmental Engineering, Grainger College of Engineering, University of Illinois at Urbana-Champaign, Urbana-Champaign, United States. [email protected].
- Department of Radiology, Mayo Clinic, Rochester, MN, USA.
- Department of Civil and Environmental Engineering, Grainger College of Engineering, University of Illinois at Urbana-Champaign, Urbana-Champaign, United States.
- Siebel School of Computing and Data Science, Grainger College of Engineering, University of Illinois at Urbana-Champaign, Urbana-Champaign, USA.
- Department of Biomedical and Translational Sciences, Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, Urbana-Champaign, USA.
Abstract
Magnetic resonance elastography (MRE) is a critical non-invasive modality for quantifying tissue stiffness, but its clinical utility is limited by the high computational cost and noise sensitivity of traditional inversion techniques, such as algebraic Helmholtz inversion (AHI) and finite element methods (FEM). This study proposes a high-performance deep learning alternative to accelerate and stabilize MRE reconstruction. We developed an open-source, physics-based simulation framework that pairs realistic liver and tumor geometries from the LiTS dataset with clinically informed stiffness priors to generate a large-scale benchmark dataset. To improve reconstruction accuracy, we introduce the U-Net Transformer (UNT), a novel architecture that hybridizes the local inductive bias of convolutional neural networks with the global context-awareness of Transformers. Numerical experiments indicate that data-driven training consistently outperforms physics-informed training in prediction accuracy. The proposed UNT architecture demonstrates significantly higher reconstruction accuracy than current state-of-the-art machine learning models. Additionally, we validated a physics-informed fine-tuning strategy that adapts the model to "sparse-data" clinical scenarios, enhancing patient-specific accuracy with minimal computational overhead. The UNT architecture and the accompanying simulation framework establish a standardized foundation for AI-driven MRE inversion. By providing a rigorous performance baseline for neural operators, this work offers a validated pathway toward real-time, AI-assisted clinical decision support in radiological diagnostics.