QSM algorithm selection for longitudinal iron tracking

A longitudinal QSM study can fail without producing an obviously bad image. The susceptibility maps may look clean, the anatomical registration may appear convincing, and the group-level statistics may even reach significance, while the measured change in deep gray matter reflects reconstruction variability more strongly than biology.

This is the central difficulty behind qsm reconstruction algorithms for longitudinal brain iron: the scan does not directly display iron concentration, and the final map is not determined by the gradient-echo acquisition alone. It is the result of a chain of processing decisions—phase unwrapping, background field removal, dipole inversion, regularization, denoising, registration, and, in some cases, physiological noise correction. Each step changes the relationship between the observed phase signal and the susceptibility value eventually entered into a clinical trial model.

For researchers following subtle changes across months or years, that relationship matters more than visual sharpness. A biomarker must be sensitive to the pathological trajectory while remaining stable when the underlying tissue has not changed. Consider the implications for studies of neurodegeneration, multiple sclerosis, or other conditions in which iron-related alterations may be gradual, spatially heterogeneous, and entangled with changes in myelin, calcium, inflammation, or tissue composition. The algorithm is not a neutral translator. It is part of the measurement.

The challenge of longitudinal reproducibility in QSM

Quantitative susceptibility mapping estimates the magnetic susceptibility of tissue from phase information acquired with gradient-echo MRI. The physical problem is ill-posed because the local field generated by tissue susceptibility is spatially filtered by the dipole kernel. In practical terms, several different susceptibility distributions can produce similar measured field patterns, particularly in orientations where the dipole response approaches its so-called magic-angle null.

Dipole inversion therefore requires a specialized reconstruction algorithm. Common approaches include MEDI, iLSQR, TGV, HEIDI, STAR-QSM, and TKD, each imposing a different set of assumptions about noise, edges, sparsity, smoothness, or the relationship between the magnitude image and the susceptibility map.

That choice becomes especially consequential in a longitudinal design. In a cross-sectional analysis, an algorithm may introduce a consistent bias that is shared across participants and therefore partly absorbed into the comparison. In a repeated-measures study, the more dangerous problem is inconsistent bias: a difference between visits that arises from motion, signal-to-noise variation, respiration, scanner drift, reconstruction settings, or the interaction between those factors and the inversion method.

A small apparent change in susceptibility can then be interpreted as increasing iron burden when it is, in fact, a change in the processing response to the same tissue.

The clinical relevance is not abstract. Many candidate QSM endpoints are expected to move slowly. Iron accumulation in structures such as the basal ganglia or substantia nigra may be biologically meaningful over a longer interval than the variability of an individual scan. If the technical noise is larger than the expected biological effect, the trial may need more participants, more visits, or a less ambitious endpoint. In the worst case, the study may identify a statistically reproducible processing artifact rather than a reproducible tissue signal.

In longitudinal QSM, reproducibility is not a cosmetic property of the map; it determines whether a measured change can plausibly belong to the patient rather than to the pipeline.

This is why pipeline standardization should begin before the first participant is scanned. The protocol, field strength, echo sampling, coil configuration, phase handling, background field removal, inversion algorithm, regularization, spatial normalization, and region-of-interest definition all form one measurement system. Changing any of them midway through a study can alter the scale and texture of the biomarker, even when the acquisition appears nominally unchanged.

There is also a biological qualification that should remain visible throughout the analysis. QSM is sensitive to tissue magnetic susceptibility, not to iron alone. Paramagnetic iron contributes to the measured signal, but diamagnetic myelin and calcium can also affect bulk susceptibility. A longitudinal increase in a region-of-interest value may therefore reflect several tissue processes, depending on anatomy and disease context. The most defensible interpretation is usually not that QSM has directly measured iron, but that it has measured a susceptibility alteration that may be biologically related to iron among other contributors.

What the evaluation of 378 pipelines tells us

A systematic evaluation of 378 QSM processing pipelines offers an unusually clear warning against treating reconstruction as a minor technical detail. The combinations varied in background field removal and dipole inversion, and their performance differed substantially in both sensitivity to deep gray matter susceptibility changes and reproducibility error.

The strongest overall performance was associated with pipelines using RESHARP for background field removal paired with AMP-PE, HEIDI, or LSQR inversion. This does not establish one universal pipeline for every scanner, field strength, acquisition, or disease cohort. It does, however, provide a practical direction for investigators deciding where to begin: background field removal and dipole inversion should be evaluated as a coupled system, rather than selected independently from a list of familiar algorithm names.

RESHARP attempts to remove slowly varying background fields while limiting the analysis to a region where the local field can be estimated reliably. That operation is not trivial near tissue boundaries. Erosion and the handling of peripheral voxels can reduce usable brain coverage, and the choice of parameters may affect structures close to the edge of the mask. A pipeline that performs well in a large central nucleus may behave differently in a thin or irregular anatomical region.

The inversion stage then determines how the remaining local field is converted into susceptibility. AMP-PE, HEIDI, and LSQR do not simply produce alternative renderings of the same information. Their regularization and numerical constraints influence noise propagation, contrast at tissue boundaries, and the degree to which ill-posed directions are stabilized. As a result, the same substantia nigra or thalamic ROI may show different apparent longitudinal sensitivity depending on the inversion method.

For a clinical trial, the useful question is therefore not merely which map looks most anatomical. It is whether a pipeline can distinguish three situations:

1. A true biological susceptibility change occurring within the target tissue.

2. A stable baseline difference between participants that does not progress.

3. A visit-specific fluctuation caused by acquisition or reconstruction variability.

Only the first should drive a longitudinal endpoint. The second can often be handled through modeling or normalization, while the third must be reduced, characterized, or incorporated into uncertainty estimates.

A practical comparison should include more than a single correlation coefficient. The analysis team should examine:

• Test–retest agreement within the same subject, preferably using the exact processing chain intended for the trial.
• ROI-specific bias, because a method can be stable in the putamen and less stable in the substantia nigra.
• Sensitivity to plausible changes in mask definition, registration, and erosion.
• Preservation of lesion or iron-related contrast after denoising and regularization.
• The degree to which susceptibility values shift when background field removal parameters are altered.
• Performance at the field strength and acquisition geometry actually used in the study.
• Whether the algorithm can be deployed consistently across sites and software environments.

This is where qsm inversion algorithm comparison becomes clinically useful rather than merely technical. The comparison should be designed around the intended inference. A method optimized for visually sharp anatomical boundaries may not be the method that best supports a small change estimate over repeated visits. Conversely, a heavily regularized map may appear impressively stable because it has suppressed the very variation that the trial hopes to detect.

Background field removal and dipole inversion should be selected together

The phrase “QSM pipeline” can conceal an important distinction. Background field removal and dipole inversion solve different problems, but errors introduced in the first stage become part of the input to the second. The two stages should therefore be tuned and validated together.

Background fields arise from sources outside the tissue region of interest, including air–tissue susceptibility interfaces and other large-scale magnetic field variations. If they are not adequately removed, the local field estimate carries spatial trends that can be misinterpreted during inversion. If the removal is too aggressive, particularly near boundaries, genuine local information may be discarded.

RESHARP-based processing has shown favorable performance in the large pipeline evaluation, particularly when paired with AMP-PE, HEIDI, or LSQR inversion. That result supports its use as a candidate starting point, not as a substitute for site-specific validation. A 3T research protocol with high-quality phase data is not interchangeable with a lower signal-to-noise acquisition, and an algorithm’s behavior in healthy adults cannot automatically be transferred to patients with lesions, atrophy, altered tissue composition, or more severe motion.

The familiar debate around MEDI vs TKD QSM accuracy also needs to be framed carefully. TKD is computationally straightforward and can be attractive for rapid processing or baseline comparisons, but thresholding the dipole kernel can influence artifacts and susceptibility scaling. MEDI incorporates magnitude information and regularization, which may improve the plausibility of the reconstructed map and, under some conditions, support strong repeatability. Yet no inversion method is intrinsically accurate in every anatomical and clinical setting.

An evaluation of nonlinear morphology-enabled MEDI with L1 regularization at 1.5T and 3T found excellent ROI reproducibility in repeated scans. Correlation coefficients were at least 0.97, with mean inter-scan differences under 1.24 parts per billion in healthy adults and under 4.15 parts per billion in patients with multiple sclerosis. These figures are valuable because they show that highly repeatable ROI measurements are achievable, including in a clinical cohort. They should not be read as a universal performance guarantee, however. Repeatability depends on the acquisition, population, registration, ROI construction, and physiological conditions, as well as on the reconstruction itself.

The 95% limits of agreement in the reported repeated-scan assessment extended from −25.5 to 25.0 parts per billion for ROI-based QSM. That range is a reminder that an excellent correlation does not mean that every individual change is measured with negligible error. Correlation describes how consistently subjects rank across visits; agreement describes how far repeated measurements may differ in absolute terms. A biomarker intended for individual monitoring needs both perspectives.

A sensible validation sequence is therefore:

1. Lock the acquisition and preprocessing conditions. Use the same echo structure, phase convention, spatial resolution, field strength, and anatomical coverage whenever possible.

2. Define candidate reconstruction families. Include at least one method with a distinct regularization philosophy rather than comparing only parameter variants within a single algorithm.

3. Run repeatability analysis in the intended population. Healthy volunteers may establish technical behavior, but patients reveal the effects of lesions, atrophy, motion, and altered tissue contrast.

4. Measure stability in the actual ROIs. Whole-brain summary metrics can hide poor performance in small nuclei.

5. Stress-test the pipeline. Perturb masks, registration, denoising, and physiological correction to learn which decisions dominate the variance.

6. Freeze the processing version before the main longitudinal analysis. If an update becomes necessary, preserve the old and new outputs and quantify the resulting shift.

This process is slower than selecting an algorithm from a software repository, but it is aligned with the way clinical evidence is built. The endpoint must be defensible not only as an image-processing result, but as a measurement of change over time.

Physiological noise is part of the biomarker problem

Respiration-induced magnetic field variations are easy to underestimate because they may not present as obvious image corruption. Breathing changes the magnetic environment around the chest and upper abdomen, and those field variations can propagate into phase-based measurements. In a technique that derives susceptibility from phase, a physiologically driven field fluctuation may appear as a change in tissue property if it is not addressed.

Correction of respiration-related magnetic field variation has been shown to improve QSM inter-scan reproducibility, with ROI measurement reproducibility improving by 5.50% in deep gray matter and 11.89% in white matter. The larger improvement in white matter is a useful reminder that physiological noise does not affect all anatomical compartments equally. A correction strategy that seems unnecessary when assessing one nucleus may be consequential for distributed white matter analyses or lesion-adjacent measurements.

The clinical implication is straightforward: respiratory monitoring or correction should not be treated as an optional refinement reserved for exceptionally rigorous imaging laboratories. If the study is designed to detect gradual susceptibility changes, physiological variation belongs in the error budget from the beginning.

This does not mean that every QSM protocol must become operationally complex. It means that the research team should determine whether respiration is recorded, how the signal is synchronized with acquisition, and whether the correction is applied consistently at every visit. A correction available only for a subset of participants can create a new source of imbalance. Likewise, adding respiratory correction after the baseline phase may improve the maps while compromising comparability with earlier scans.

The same principle applies to motion and scanner-related variation. In longitudinal neuroimaging, a participant’s trajectory is partly a biological trajectory and partly a trajectory through changing acquisition conditions. If the study spans multiple scanners or software versions, the analysis should preserve those details rather than treating all measurements as if they came from one stable instrument.

The quieter the expected biological change, the more explicitly the study must characterize breathing, motion, registration, and reconstruction as competing explanations for the observed signal.

From a single map to a longitudinal endpoint

A QSM map becomes a biomarker only after it is connected to a defined biological and statistical question. The endpoint might be a mean susceptibility value in a deep gray matter nucleus, a spatially resolved pattern across several structures, a lesion-specific measure, or a rate of change estimated from multiple visits. Each choice places different demands on reconstruction.

For a mean ROI value, local noise may average out, but registration and segmentation errors can still dominate in small structures. For a voxelwise analysis, spatial correspondence and regularization become more consequential. For a slope across visits, the number and spacing of scans influence the distinction between a stable trajectory and a transient fluctuation. There is no single notion of reproducibility that answers all of these questions.

The expected biological effect should guide the technical design. If the study aims to detect a slow increase in susceptibility over a relatively long interval, a pipeline with slightly lower apparent contrast but stronger inter-scan stability may be more useful than a visually aggressive reconstruction. If the endpoint concerns focal lesions, preserving local variation may take priority, but that sensitivity must be demonstrated rather than assumed.

This is also where simultaneous multi-time-point reconstruction offers an important conceptual shift. Rather than reconstructing each visit independently, a longitudinal QSM framework jointly reconstructs susceptibility maps across multiple time points while enforcing temporal spatial sparsity. The aim is to reduce inter-scan variability by using information about the series itself, while preserving genuine lesion and iron-related changes.

Joint reconstruction is attractive because longitudinal data contain structure that single-visit processing cannot use. A stable region should not fluctuate arbitrarily from one scan to the next, while a true change should remain possible where the data support it. Temporal regularization can therefore act as a form of prior knowledge about biological continuity.

But that prior must be handled with care. If temporal smoothness is too strong, a real abrupt change may be attenuated. If spatial sparsity is imposed without regard to anatomy, diffuse or subtle alterations may be underrepresented. A framework designed to improve consistency can, if poorly calibrated, make the trajectory appear more orderly than the biology actually is.

For this reason, simultaneous reconstruction should be evaluated against independently reconstructed time points, simulated changes, and repeated scans. The key question is not whether the joint method reduces variability. It is whether it reduces variability while retaining sensitivity to the magnitude, location, and timing of changes that matter clinically.

A useful reporting framework for a longitudinal QSM study could include the following elements:

The study should also distinguish technical repeatability from biological responsiveness. A map can be highly repeatable in healthy adults and still fail to track disease progression. Conversely, a patient cohort may show genuine heterogeneity that lowers simple agreement statistics without invalidating the endpoint. The appropriate analysis depends on whether the intended use is group-level trial inference, individual monitoring, or both.

Standardization without pretending that one pipeline fits all

The phrase “quantitative susceptibility mapping pipeline standardization” can suggest that the field is waiting for one definitive reconstruction recipe. Clinical translation is unlikely to be that simple. Standardization is better understood as making the entire measurement process explicit, versioned, testable, and comparable across time and sites.

At a minimum, a longitudinal protocol should document:

• The magnetic field strength and scanner platform.
• Echo times, echo spacing, flip angle, resolution, and acquisition orientation.
• Phase preprocessing, unwrapping, and sign conventions.
• Brain masking and erosion rules.
• Background field removal method and parameters.
• Dipole inversion algorithm, regularization, and stopping criteria.
• Treatment of magnitude information and low-signal regions.
• Registration strategy and interpolation method.
• ROI or atlas definition, including how small nuclei are handled.
• Physiological noise correction, motion handling, and quality exclusions.
• Software version, computational environment, and any parameter changes over time.

This level of documentation is not bureaucratic overhead. It allows a future analyst to determine whether an apparent change in susceptibility is compatible with the intended biological hypothesis or whether it coincides with a processing transition.

A harmonized protocol should still permit local validation. The best performing combination in a large pipeline evaluation is a valuable candidate, but an endpoint intended for oncology MRI assessment, neurodegeneration, or multiple sclerosis may have different demands. Lesion burden, atrophy, susceptibility sources, and motion patterns alter the problem. Field strength also matters, and the universally optimal parameters for all magnetic field strengths have not been established.

The same caution applies to short time frames. There is no established universal threshold for detecting mild microstructural iron shifts in small subregions over intervals shorter than six months. A study that expects to resolve such changes should not infer sensitivity from visual map quality alone. It needs repeated acquisitions, explicit uncertainty estimates, and a biological rationale for why the expected change should be detectable within that interval.

What a defensible QSM selection decision looks like

A reconstruction choice is defensible when it can answer three questions in the language of the intended trial.

First, what tissue property is the map being used to represent? If the endpoint is described as brain iron, the protocol should state why susceptibility is considered a useful proxy in the selected structure and how calcium, myelin, or other contributors will be handled in interpretation.

Second, what size and pattern of change must the pipeline detect? A method can be suitable for large group differences and unsuitable for subtle individual trajectories. The expected effect should determine the repeatability target, the ROI design, and the number of visits.

Third, what technical evidence shows that the observed change is not merely a reconstruction effect? This is where the combination of test–retest data, physiological correction, parameter sensitivity analysis, and locked software becomes more persuasive than any isolated algorithm label.

For many studies, RESHARP paired with AMP-PE, HEIDI, or LSQR provides a strong set of candidates because these combinations performed well in the systematic assessment of longitudinal sensitivity and reproducibility. L1-regularized MEDI also demonstrates that excellent ROI repeatability is achievable at both 1.5T and 3T, including in a multiple sclerosis cohort. These findings narrow the search, but they do not eliminate the need for local validation.

TKD may remain useful as a baseline or computationally accessible comparison, particularly when a study needs to understand how thresholding affects the result. However, it should not be selected on convenience alone when the endpoint depends on small longitudinal changes. The relevant comparison is not which algorithm is easiest to run, but which one preserves a trustworthy relationship between tissue biology and measured change under the conditions of the study.

The most mature workflow may ultimately be a two-stage design: an initial technical qualification phase using repeated scans and candidate pipelines, followed by a locked clinical analysis in which the reconstruction is treated as part of the endpoint definition. Simultaneous multi-time-point reconstruction can then be assessed as a prespecified alternative or sensitivity analysis rather than introduced after results begin to emerge.

That separation protects the biological interpretation. It prevents the pipeline from becoming an adjustable instrument tuned to produce a desired trajectory, and it gives readers, regulators, and collaborators a clearer view of what the biomarker can—and cannot—support.

The trajectory matters more than the map

QSM is often introduced through its visual promise: a map that renders susceptibility differences across the brain with a level of contrast unavailable to conventional magnitude imaging. For clinical research, its deeper value lies elsewhere. It may allow investigators to follow a tissue process over time, provided the measured trajectory is more stable than the technical uncertainty surrounding it.

This shift allows us to view algorithm selection as part of translational neuroscience rather than as an afterthought in image processing. Background field removal, dipole inversion, physiological correction, registration, and temporal regularization all influence whether a subtle change can be connected to pathology, treatment response, or cognitive reserve.

The field does not need a single algorithm that wins every comparison. It needs pipelines whose behavior is understood in the populations, scanners, structures, and time scales where they will be used. It needs honest separation between susceptibility and iron, between correlation and agreement, and between reduced noise and preserved biological sensitivity.

A well-designed longitudinal QSM study therefore begins with restraint. It does not promise that one biomarker will explain disease, or that a cleaner map will resolve an uncertain mechanism. It asks a narrower and more useful question: can this reconstruction measure a biologically meaningful change repeatedly enough to support a clinical inference?

When the answer is supported by test–retest evidence, physiological control, transparent standardization, and a pipeline chosen for the endpoint rather than for its reputation, QSM becomes more than an attractive image. It becomes a carefully qualified way of observing how human tissue changes through time.