HSSRecon partitions data by physical acquisition units and learns a Hankel subspace, achieving competitive reconstruction metrics across six aggregated fastMRI brain conditions
Synopsis
The work proposes HSSRecon, a scan-specific self-supervised reconstruction framework for parallel MRI: it partitions data by physical acquisition units before Hankel lifting and applies multiplicity normalization to repeated Hankel copies in overlapping windows, the network learns a compact complex-valued Hankel subspace operator, and reconstruction is performed over the original k-space variables with a conjugate-gradient solver and hard data consistency; on fastMRI brain data with three contrasts and three sampling masks across six aggregated conditions, it achieves competitive PSNR, SSIM, and NMSE.
in Fig. 1(a).
arXiv · Page 2Interpretation
HSSRecon separates structural learning in the Hankel domain from data consistency in the physical domain: instead of learning a mapping that directly predicts missing data, the network learns a compact complex-valued Hankel subspace operator, and reconstruction is performed over the original k-space variables using a conjugate-gradient solver with hard data consistency. Relative to reconstruction that directly learns a missing-data mapping, this design separates the exploitation of multi-coil and local Hankel correlations from solving over unacquired degrees of freedom. The design is supported by the method description and theoretical analyses, and evaluated on fastMRI brain data across three contrasts and three sampling masks in six aggregated conditions.
To address data leakage in Hankel self-supervision, HSSRecon partitions data by physical acquisition units before Hankel lifting and applies multiplicity normalization to repeated Hankel copies in overlapping windows. This responds to the risk that splitting lifted Hankel entries for self-supervision can place the original sample in both input and target. The text provides theoretical analyses of physical-group splitting and multiplicity normalization.
The work establishes positive definiteness, uniqueness, hard data consistency, and a finite-step conjugate-gradient error bound for the system. These properties connect structural learning in the Hankel domain with the data-consistency solve in the physical domain, giving a theoretical characterization of solver behavior. The conclusions come from theoretical analyses in the text, not from experimental observation alone.
On fastMRI brain data, HSSRecon achieves competitive peak signal-to-noise ratio, structural similarity, and normalized mean squared error across six aggregated conditions. The evaluation covers three contrasts and three sampling masks, indicating the framework operates across multiple acquisition settings. The evidence is six aggregated conditions on fastMRI brain data; no specific numerical values are given in the text.
Perspective
The result targets scan-specific reconstruction for parallel MRI, applies to multi-coil acquisition under undersampling, and is evaluated on fastMRI brain data with three contrasts and three sampling masks across six aggregated conditions. For researchers and implementers interested in separating Hankel-domain structural learning from physical-domain data consistency, the framework offers a reusable design; its theoretical conclusions concern the constructed system, and its experimental conclusions are limited to the evaluated scope.
The text does not give specific values for peak signal-to-noise ratio, structural similarity, or normalized mean squared error, nor itemized comparisons against baselines, so the magnitude of the advantage cannot be judged from the summary. How multiplicity normalization and physical-group splitting behave under different sampling masks, contrasts, or acceleration factors, and how tight the finite-step conjugate-gradient error bound is in practice, remain to be examined with the full text and further settings.