Decoding the Phase Signal in Magnetic Resonance Imaging
Every MRI signal has two parts: magnitude and phase.
The magnitude tells us how strong the signal is—this is what creates the familiar black-and-white MRI images showing the shape and size of tissues.
The phase tells us something deeper—it captures subtle differences in how tissues interact with the magnetic field.
Both are part of the full picture. Measurements using the magnitude of the reconstructed MR image has been the predominant practice in various clinical applications, as it a rapid way to visualize results.
Recently, the phase of the MR signal has been shown to be more sensitive and precise than magnitude at measuring indications of tissue disease such as: the electromagnetic properties of tissues (susceptibility and conductivity), iron deposition concentration, calcification, tissue temperature, venous oxygenation, blood velocity, tissue elasticity, among others. The MR phase has also been extensively used for “fieldmap estimation” in the neuroimaging community to calculate and correct for imperfections of the magnetic field.
In phase-sensitive MRI sequences, the magnetic susceptibility differences between various tissues, such as hematoma and surrounding brain tissue, can lead to signal phase shifts. These shifts can be visualized and quantified, enhancing the contrast between regions affected by microbleeds and healthy brain tissue. As a result, phase imaging provides a more detailed depiction of microhemorrhages, potentially leading to earlier diagnosis and better understanding of their implications in conditions such as Alzheimer’s disease and cerebral amyloid angiopathy.
Despite all this, all MRI protocols and post-processing algorithms are optimized for magnitude-domain contrast or signal-to-noise ratio (SNR). One may thus ask, how can we optimize an MRI protocol if the information of interest resides in the phase domain? It turns out that SNR-optimal MR phase acquisition and estimation is a bit more difficult than magnitude. Here is why:
There are three phase-imaging operation regimes that MR protocols can choose from:
(I) The noisy phase: here, the image is rapidly acquired before significant decay of the magnitude image, yet before sufficient accumulation of a phase signal: this regime is dominated by phase-noise, or
(II) The noisy magnitude: here, the phase signal is allowed to fully accumulate, yet at a time when most of the magnitude of the signal has significantly decayed: this regime is dominated with magnitude-noise which impacts the phase measurement, or
(III) The ambiguous phase: this regime between falls between regimes I & II. Here, the phase signal accumulates enough, without too much decay in magnitude. In this regime, a phase wrapping ambiguity dominates. Because the phase is measured on a circle, allowing the signal to accumulate more than 360 degrees introduces ambiguities about the true value of the underlying frequency (phase = 2x pi x frequency x time). This regime also suffers from errors related to noise (both phase and magnitude induced), as is common with any measurement. So the challenge in this regime would be to disambiguate phase errors from both phase-wrapping and noise contributions.
The figure above shows a schematic that visually explains these challenges above. What complicates matters more is that it is not always possible to choose which regime to operate in due to hardware constraints (minimum echo time spacing, gradient strengths, bandwidth) or total imaging time restrictions.
We have presented a solution to this dilemma in a series of papers published in MRM, NMR in Biomedicine, AJNR and others (attached below).
In the MRM papers, we show that it is mathematically possible to optimally estimate the phase of the MR signal using a multi-echo protocol, without increasing the total acquisition time. Our solution, coined MAGPI, frames phase estimation as a maximum likelihood (ML) optimization problem. A key innovation is the joint design of the acquisition protocol and reconstruction algorithm so as to maximize the phase SNR, even in low magnitude SNR regimes.
MAGPI does not suffer from any of the tradeoffs of the phase estimation methods (phase wrapping, noise and unknown phase offsets, outlined in figure above). The resulting maximum SNR phase image does not contain any of the artifacts seen in traditional phase images.
The framework enables shorter acquisition times, improved spatial resolution, and better contrast-to-noise ratio (CNR) in both phantom and clinical brain imaging settings.
In clinical studies, MAGPI produced diagnostic-quality brain phase maps in under 2.5 minutes at sub-millimeter resolution, and achieved factors of gains in contrast-to-noise ratio even in low-SNR regimes.
The phase of the MRI signal plays a crucial role in detecting microbleeds in the brain. Microbleeds, which are small hemorrhagic events often associated with small vessel disease and other neurological conditions, can be challenging to identify using standard MRI techniques. However, phase imaging offers unique advantages due to its ability to exploit the magnetic susceptibility effects of blood oxygenation levels. Our proposed method expands the clinical and research utility of phase imaging, especially in susceptibility-weighted imaging and other applications requiring fine structural detail at high resolution.
As an example, one clinical study evaluates the performance of MAGPI-SWI, a phase imaging technique derived from the MAGPI framework, for detecting traumatic cerebral microbleeds (CMBs) in athletes with mild traumatic brain injury (mTBI). Unlike conventional susceptibility-weighted imaging (SWI), which suffers from phase noise, susceptibility artifacts, and limited signal-to-noise ratio (SNR) near the skull base, MAGPI-SWI leverages a mathematically rigorous phase reconstruction method that offers higher phase SNR and reduced susceptibility distortions. The study involved 10 collegiate athletes who were scanned at 2 days, 2 weeks, and 6 weeks post-injury using both standard SWI and MAGPI-SWI protocols. A neuroradiologist, blinded to the scan type, evaluated images for CMBs and compared image quality. MAGPI-SWI identified CMBs in 6 participants—including all 4 with concurrent contusions on FLAIR—while traditional SWI failed to detect these clearly due to artifact interference.
The improved image quality in MAGPI-SWI was attributed to superior phase estimation, optimized echo-time selection, and more effective high-pass filtering. These enhancements enabled MAGPI-SWI to detect subtle microbleeds, particularly in artifact-prone areas like the inferior frontal lobe. Furthermore, MAGPI-SWI produced higher contrast-to-noise and delineated deep brain structures more clearly than traditional methods. The findings suggest that MAGPI-SWI significantly improves the detection of small hemorrhagic lesions that are often missed with conventional SWI, and could become a valuable tool for early and accurate diagnosis of mTBI. The study also emphasized that these gains stem from increased phase SNR and efficient data integration, not just better unwrapping, making MAGPI-SWI a fundamentally more robust imaging technique.
In quantitative imaging, MAGPI has been evaluated in an independent study that systematically compares a wide range of phase estimation methods for Quantitative Susceptibility Mapping (QSM).
Using a precisely controlled rotating-tube phantom designed to simulate biologically relevant susceptibility values. The authors evaluated 90 combinations of MRI acquisition sequences (single-echo GRE, multiecho GRE, sEPI, and MAGPI) and ten different phase estimation techniques, including Laplacian, region-growing, branch-cut, temporal methods, and the maximum-likelihood-based MAGPI algorithm. The phantom setup allowed for ground-truth comparison using known susceptibility differences and analytical modeling, enabling a rigorous assessment of each method’s accuracy, repeatability, and robustness. The study focused specifically on Steps 1 and 2 of the QSM pipeline: acquisition and phase estimation—where early errors can propagate and compromise downstream analysis.
The results highlight substantial variability in performance across methods. Many traditional unwrapping approaches were prone to large errors, especially in low-SNR or high-susceptibility regions. In contrast, MAGPI and its unoptimized variant (MAGPI-unopt) consistently achieved the highest accuracy and reliability across all scan conditions and tube contents, with over 90% of relative phase errors falling below 10%. Laplacian-based methods yielded smooth-looking maps but were often quantitatively inaccurate. The study underscores that robust, user-independent phase estimation—especially using likelihood-based approaches like MAGPI—is crucial for reliable QSM, and that seemingly subtle differences in acquisition or post-processing choices can have a significant impact on clinical viability.