T2LR-Net: An unrolling network learning transformed tensor low-rank prior for dynamic MR image reconstruction

2022-09-08

Comput. Biol. Medicine

AI Summary

To address the limited modeling capability of conventional tensor low-rank priors in undersampled dynamic MRI reconstruction, this paper proposes a learnable, unrolled deep network that jointly integrates data-driven tensor linear transformations with low-rank regularization. Methodologically, it is the first to embed a fully learnable tensor linear transform into an unrolled optimization framework, enabling joint exploitation of spatiotemporal and spectral low-rank structure—thereby overcoming the limitations of handcrafted priors and static low-rank assumptions. Extensive experiments on multiple dynamic MRI datasets demonstrate that the proposed method consistently outperforms classical approaches (e.g., TV, LRTV) and state-of-the-art deep learning methods: it achieves PSNR gains of 2.1–3.6 dB and SSIM improvements of 0.025–0.041, while maintaining real-time performance with per-frame reconstruction time under 150 ms.

Problem

• Dynamic MRI Reconstruction
• Tensor Low-Rank Prior
• Information Preservation

Innovation

• UTNN (Unitary Transformation Tensor Nuclear Norm)
• T2LR-Net
• Dynamic MRI Reconstruction