Preserving MRI’s hidden phase signal could improve accelerated brain imaging

Why brightness alone is not enough

Multi-echo gradient-echo MRI records signals at several times after excitation. How quickly their strength fades helps estimate T2*, a measure of signal decay. Changes in phase support quantitative susceptibility mapping, or QSM, which estimates how tissue responds to a magnetic field.

Collecting fewer measurements can accelerate imaging, but leaves reconstruction software with a harder problem. Later echoes have weaker signals, while phase can change rapidly across nearby locations, especially at high magnetic fields. A reconstruction algorithm may struggle to distinguish those genuine variations from artifacts caused by missing measurements.

Juan Liu and Nian Wang’s proposed method, MSPR-GRE, addresses that problem by updating its phase assumptions as the reconstruction improves, rather than relying on one initial estimate.

Reconstruct, update phase, repeat

The framework works from coarse information toward finer detail. Its first stage uses central measurements from k-space—the MRI measurement domain—to reconstruct an initial image and estimate phase. Later stages incorporate measurements covering finer spatial detail.

Before the second stage, the method uses earlier, more reliable echoes to fit phase evolution when at least three echoes are available. That fit supplies an updated phase estimate for the reconstruction; it is not treated as an independently validated magnetic-field map. The third stage takes its phase estimate directly from the second-stage images.

Each stage alternates a learned image-refinement step with a physics-based step that checks agreement with measured data. For acquisitions with at least four echoes, later stages also represent signal evolution using a compact set of simulated patterns. In the eight-echo mouse experiment, this reduced the number of reconstructed coefficient maps from eight to three. Shorter echo trains are reconstructed directly without that compression.

Training is self-supervised: some acquired measurements are withheld from reconstruction and used to score its predictions. Crucially, those withheld samples are also excluded from phase estimation and initialization. The stages are trained sequentially, freezing earlier stages. At inference, all acquired measurements become available for reconstruction.

What the experiments showed

The researchers evaluated two high-resolution ex vivo mouse-brain datasets acquired at 9.4 tesla, alongside human brain data from the 7-tesla AHEAD database. The mouse acquisitions used eight echoes at 30-micrometre resolution and two echoes at 18-micrometre resolution.

For AHEAD, ten datasets were used for training and five independent datasets for testing. The researchers retrospectively removed measurements at nominal acceleration factors of four, eight and twelve.

At the highest factor, fourth-echo magnitude-image error was 12.09% ± 0.60% with MSPR-GRE, compared with 21.55% ± 1.12% for SSDU, an established self-supervised reconstruction approach, and 16.67% ± 0.76% for wavelet regularization. These figures are normalized root mean square errors: lower values indicate closer agreement with the reference. The reported variation is the standard deviation across five test datasets.

Derived tissue maps also improved relative to SSDU. At the same acceleration factor, T2* map error fell from 74.11% to 52.00%, while QSM error fell from 61.98% to 50.10%. Those remaining errors matter: better agreement is not equivalent to accurate clinical measurement.

The authors report fewer visible artifacts and better preservation of fine anatomy in the mouse images. Removing phase-related components weakened results in qualitative comparisons, supporting their proposed mechanism. Most gains from adding stages occurred between one and two stages; a fourth stage brought little consistent benefit in the tested eight-echo dataset.

What remains unproven

The human acceleration factors refer to a processed, released Cartesian k-space grid, not fully sampled raw scanner acquisitions. The original scans already used parallel imaging. Consequently, a nominal factor of twelve here does not demonstrate a twelvefold reduction in clinical scan duration. Human evaluation was retrospective, although the mouse experiments included a prospectively undersampled acquisition.

The evidence base is small. Mouse reference images retained noise and possible phase inconsistencies, and MSPR-GRE did not win every metric against conventional methods. Phase accuracy itself was not directly quantified. Results therefore leave open how reliably the approach transfers across scanners, acquisition protocols and more varied subjects.

The November 9 arXiv preprint is not established peer-reviewed clinical evidence. It supports further investigation of progressive phase refinement—not claims of improved diagnosis. The authors state that code will be released upon publication, leaving independent reproduction as an important next step.