Multi-contrast MRI acceleration via post-reconstruction fusion.
Authors
Affiliations (3)
Affiliations (3)
- School of Electrical and Computer Engineering, Tel Aviv University, Tel Aviv, 6997801, Israel. Electronic address: [email protected].
- School of Electrical and Computer Engineering, Tel Aviv University, Tel Aviv, 6997801, Israel.
- Diagnostic Imaging, Sheba Medical Center, affiliated with the Gray School of Medicine, Tel Aviv University, Tel Aviv, Israel.
Abstract
Magnetic Resonance Imaging (MRI) is the gold standard for neuroimaging, yet routine brain protocols require multiple high-resolution 3D contrasts (e.g., T1, T2, and T2-FLAIR), resulting in long scan times. Many deep-learning acceleration methods assume access to raw k-space and rely on non-Cartesian trajectories or pseudo-random undersampling patterns, which can require specialized sequence implementations. In this work, we present a practical protocol-level acceleration framework enabled by complementary orthogonal Cartesian acquisitions that can be executed using standard scanner settings. Specifically, each contrast is acquired with reduced phase-encoding matrix size along a contrast-specific axis (a fourfold reduction along one contrast-specific phase-encoding dimension), producing rapidly acquired volumes with axis-specific resolution loss but complementary spatial-frequency content across contrasts. We propose the Frequency Attention Residual Denoising (FARD) network, a multi-contrast fusion model that leverages both spatial and frequency-domain processing to enhance apparent isotropic detail on a 1mm<sup>3</sup> grid for all contrasts from these complementary inputs. We evaluate the approach using two complementary settings: a prospectively acquired clinical-scanner dataset that reflects real acquisition and vendor-reconstruction conditions, and a controlled retrospective simulation on BraTS-GLI 2024, which provides large-scale pathology-containing multi-contrast data but does not model prospective scanner effects. On the prospective cohort, the proposed framework achieves approximately 4× protocol acceleration (10.4 to 2.6 min), provides the strongest or near-strongest PSNR and SSIM among the evaluated reference-free methods, and approaches reference-guided performance without requiring a high-resolution reference contrast. BraTS experiments provide complementary large-scale evidence of algorithmic effectiveness under controlled retrospective degradation.