Undersampled K-space information recovery with long-range temporal memory for multi-coil MRI reconstruction.
Authors
Affiliations (2)
Affiliations (2)
- Computer Engineering Department, Istanbul Technical University, Istanbul, Turkey. Electronic address: [email protected].
- Computer Engineering Department, Istanbul Technical University, Istanbul, Turkey. Electronic address: [email protected].
Abstract
Accelerated magnetic resonance imaging through k-space undersampling introduces a challenging inverse problem: recovering the full frequency spectrum from incomplete Fourier measurements. Existing unrolled networks apply static regularization uniformly across cascades, ignoring that reconstruction challenges evolve from coarse artifact suppression to fine detail recovery. We aimed to develop a temporal memory framework that accumulates reconstruction knowledge and dynamically adapts regularization strategies. Our main contribution is a cascade-wise attention-weighted memory that aggregates a buffer of recent reconstruction features through a learned, state-dependent gate, so that the regularization adapts across the unrolled cascades. We evaluated our framework on cardiac data (300 cases, 1.5T/3.0T/5.0T, nine disease categories) and knee data (197 volumes) across uniform, Gaussian, and radial undersampling at 8, 16, and 24 times acceleration. Quality was assessed using structural similarity index and peak signal-to-noise ratio. Our method achieved structural similarity of 0.888 and peak signal-to-noise ratio of 33.29 dB on cardiac data, outperforming PromptMR by 0.023 SSIM and 1.51 dB and also improving on HierAdaptMR, and ranking among the top teams in the CMRxRecon2025 challenge. On knee data at 24 times acceleration, improvements reached 0.094 for uniform and 0.127 for radial undersampling. Performance gains increased with reconstruction difficulty. Long-range temporal memory enables cascade-adaptive k-space recovery, providing substantial benefits precisely when reconstruction becomes severely ill-posed, with consistent improvements across two anatomies and several undersampling patterns, using a separately trained model for each anatomy. Code available at https://github.com/Ruru-Xu/cmr-temporal-memory.