In 2022, a university imaging center replaced its aging 1.5-tesla MRI scanner with a new 3-tesla model. The upgrade was routine — better signal, faster acquisitions, more contrast. But when researchers compared resting-state functional connectivity maps collected before and after the swap, they found something unsettling: 13 of the 20 canonical networks had changed shape, vanished, or been replaced by new ones. Only 7 maps remained stable across the two machines. The result, published in a technical report, became a quiet alarm bell for a field that often treats its hardware as a black box.

A Field's Quiet Infrastructure Crisis

Neuroscience has poured billions into large-scale projects like the Human Connectome Project and the Adolescent Brain Cognitive Development (ABCD) study. These efforts depend on dozens of scanners running the same protocols at different sites. But scanners age, magnets drift, and when a site upgrades its machine, the data before and after the upgrade may no longer be comparable. The 13-of-20 finding is not an outlier; it is a symptom of a systemic problem.

Funding agencies rarely budget for mid-study hardware replacements, and grant reviews seldom ask about scanner history. A lab that published a resting-state map in 2018 using a 1.5-tesla machine may be unable to replicate that map in 2025 on a 3-tesla machine. The field's reproducibility crisis is partly an equipment crisis, masked by the assumption that the brain's intrinsic connectivity is stable enough to survive a magnet swap.

The scale of the problem is hard to quantify. As of late 2024, roughly 40% of academic imaging centers in the United States had upgraded at least one scanner in the previous five years, according to informal surveys. Many of those upgrades happened without pre-registered hardware change dates or systematic phantom calibration. The result is a growing archive of data whose compatibility is unknown.

Not everyone agrees that the problem is as severe as it seems. Mark T. Nelson, a neuroimaging methodologist at the University of Southern California, argues that resting-state networks are robust enough to show the same general patterns across field strengths. "The core topography of the default mode network is visible at 1.5T and 3T," he notes. "What changes is the fine-grained spatial extent and the number of components you can extract, not the fundamental organization." But the 13-of-20 result suggests otherwise. The networks that disappeared included the dorsal attention network and parts of the default mode network, both central to cognitive neuroscience. If these maps are sensitive to a change in magnet strength, then decades of work comparing across sites may need re-examination.

The Scanner Upgrade That Rewrote a Lab's Data

The university in question ran a resting-state fMRI protocol on its 1.5-tesla Siemens Avanto for years. The protocol was standard: 200 volumes, TR 2 seconds, 3-mm isotropic voxels. The lab had published several papers using those data, describing the topography of the default mode, salience, and executive control networks. Then the Avanto reached end-of-life, and the center replaced it with a 3-tesla Siemens Prisma.

Before clinical use began, the lab ran the same protocol on 30 healthy volunteers who had previously been scanned on the old machine. The same participants, the same task-free condition, the same preprocessing pipeline — except the magnet. When the team applied independent component analysis to extract 20 resting-state networks, they found that 13 components did not match between the two sessions. Some networks had merged, others split, and a few appeared only in the 3-tesla data.

The 7 networks that survived were mostly primary sensory and motor maps: the visual cortex, somatomotor areas, and the auditory network. The higher-order association networks — those involved in attention, memory, and self-referential thought — were the ones that shifted. The default mode network, for example, showed a different spatial extent and a slightly anterior shift on the 3-tesla machine.

The lab published its findings as a pre-print and later in a peer-reviewed journal, with a clear message: scanner upgrade effects can be as large as the effects of aging or disease that many studies aim to detect. The paper included a table comparing each component's spatial correlation across sessions. For the 13 unstable networks, the correlation dropped below 0.3, far below the typical threshold for claiming replication.

Why Small Hardware Changes Break Big Results

Blood-oxygen-level-dependent (BOLD) contrast depends on the magnetic field strength. At 3 teslas, the signal-to-noise ratio is roughly double that at 1.5 teslas, but the sensitivity to susceptibility artifacts also increases. Vessels of different sizes contribute differently to the BOLD response at different field strengths, altering the spatial pattern of activation.

Motion artifacts also differ. A 3-tesla scanner is more sensitive to physiological noise — breathing, cardiac pulsation, head movement — than a 1.5-tesla machine. Even with the same motion-correction algorithm, the residual artifacts can shift connectivity estimates. The result is that a network that appears as a single component on one scanner may split into two components on another, or merge with a neighboring network.

Coil geometry adds another layer. The new scanner used a 64-channel head coil instead of the old 12-channel coil. Higher channel counts improve parallel imaging but also change the spatial sensitivity profile. Voxels near the surface of the brain receive more signal relative to deep structures, altering the apparent connectivity of subcortical regions. The lab's analysis pipeline, tuned to the old coil's noise structure, may have amplified these differences.

Even the same manufacturer and field strength can produce different results if the scanner model changes. A Siemens Prisma and a GE MR750, both 3 teslas, do not produce identical resting-state maps. The 13-of-20 finding is a stark reminder that the scanner is not a neutral measuring device; it is an active part of the measurement system that shapes the data in ways that are hard to predict.

The Incentive Structure That Ignores Equipment Drift

Why does the field not track hardware changes more carefully? The answer lies in incentives. Grant reviewers focus on scientific questions, not on whether the scanner at site A was upgraded halfway through a five-year study. Journals ask authors to report the scanner model and field strength, but they rarely require machine logs or phantom data. The cost of upgrading a scanner — roughly US$ 2–3 million — is a capital expense that institutions handle separately from research grants, so there is no natural mechanism to flag the change.

Publication pressure compounds the problem. A lab that discovers its post-upgrade data do not replicate its pre-upgrade findings faces a dilemma: publish a negative result about reproducibility and risk undermining its own past work, or set the old data aside and move forward. Most choose the latter. The 13-of-20 paper is an exception, and it was only possible because the lab had the resources and institutional support to run a replication study before starting new experiments. The same dynamic plays out in large consortia: when a site upgrades its scanner, the project managers often decide to simply continue with the new data and apply harmonization algorithms post hoc, rather than conducting a costly in vivo validation.

The problem is not limited to MRI. Similar issues have surfaced in other areas of science. For example, an uncalibrated optogenetic laser can swap place cell maps in rodents (see this related story), and a single unindexed vial can warp antibiotic susceptibility curves (see this earlier report). The common thread is that small changes in measurement tools can produce large changes in outcomes, yet the reward system ignores the tools.

Some funding agencies have begun to take notice. The National Institutes of Health now requires some multi-site studies to include a data harmonization plan, but these plans often focus on post-hoc statistical corrections rather than on preventing hardware-driven variance in the first place. A better approach would be to mandate pre-registration of hardware changes and to budget for mid-study phantom calibration.

What This Means for Large-Scale Consortia

Large consortia like the Human Connectome Project (HCP) and ABCD were designed to pool data across many sites to achieve statistical power. But if each site upgrades its scanner at different times, the resulting batch effects can mimic or mask developmental and clinical effects. The ABCD study, for instance, follows children over 10 years across 21 sites. Several sites have already upgraded scanners during the study. Without careful tracking, an apparent change in brain connectivity between ages 10 and 12 could be partly an artifact of a magnet swap.

Harmonization algorithms, such as ComBat, are often used to remove site effects. But these methods assume that site differences are constant over time. If a site's scanner changes mid-study, the assumption is violated. The algorithm may overcorrect or undercorrect, introducing new biases. The 13-of-20 result shows that the magnitude of the scanner effect can exceed the typical site effect, making harmonization unreliable.

Longitudinal studies are especially vulnerable. A within-subject design that scans the same participants on the same machine should, in theory, control for hardware. But if the machine is replaced between the baseline and follow-up visits, the design breaks. The lab's pre-post comparison is a microcosm of this problem: the same participants, the same protocol, but different hardware, and the result was a 65% failure rate in network replication.

Continuous phantom calibration can help. A phantom — a stable artificial object — scanned daily or weekly produces a time series of signal-to-noise, drift, and homogeneity. If the phantom metrics change after an upgrade, the data before and after can be adjusted. But few sites publish their phantom logs, and even fewer tie them to individual studies. The infrastructure to track instrument drift exists, but the culture does not demand it.

Practical Steps to Track and Mitigate Instrument Drift

What can be done? First, every imaging center should run daily quality-assurance scans on a phantom and make those logs publicly available alongside the data they accompany. A simple CSV file with SNR, drift, and uniformity metrics would allow downstream users to flag potential breaks. This is cheap and already done in many clinical settings, but research sites often skip it to save time.

Second, pre-registration of hardware change dates should be standard. If a lab plans to upgrade its scanner, it should file a brief note — before the upgrade — specifying the old and new models, the planned transition date, and any overlap scanning of phantoms or traveling subjects. This would allow the community to assess the impact on ongoing studies.

Third, traveling-subject protocols — scanning the same volunteers on multiple machines — should be mandatory for multi-site studies at the start and after any hardware change. The cost is modest compared to the cost of collecting unusable data. The 13-of-20 study essentially did this for two scanners at one site; scaling it to 20 sites would be logistically harder but feasible.

Fourth, funding agencies should explicitly budget for mid-study equipment replacements and for the validation studies that must follow. A grant that runs for five years has a non-negligible chance of experiencing a scanner upgrade. Including a line item for phantom materials, traveling-subject scans, and a data harmonization specialist would make the project more robust. Some agencies are moving in this direction, but the pace is slow.

Finally, journals should require more detailed hardware reporting. Instead of merely listing the scanner model and field strength, authors should report the coil type, software version, and a summary of quality-assurance metrics for the period of data collection. This would not eliminate the problem but would make it visible.

Consider a concrete example from the UK Biobank, which scans participants across multiple sites using identical 3-tesla Siemens Skyra scanners. Despite careful harmonization, subtle differences in coil performance and gradient linearity have been documented between sites. If one of those scanners were to be replaced, the resulting data discontinuity could affect thousands of participants. The UK Biobank has a rigorous quality assurance program, but not all studies have such resources. Smaller labs, with a single scanner and limited funding, are especially vulnerable: a single upgrade can render years of baseline data incomparable.

Another example comes from the Alzheimer's Disease Neuroimaging Initiative (ADNI), which has been scanning participants since 2004. Over that period, many sites have upgraded from 1.5T to 3T scanners, and ADNI has implemented a series of phantom-based calibration procedures to maintain consistency. Yet even with these measures, studies have shown that field strength differences contribute to variance in hippocampal volume measurements and cortical thickness estimates. The lesson is that hardware changes are not just a nuisance for functional connectivity; they affect structural imaging as well.

The 13-of-20 story is about more than one scanner upgrade. It is a reminder that neuroscience — like all empirical science — depends on instruments that are imperfect, aging, and changing. The field has focused heavily on software: preprocessing pipelines, statistical models, machine learning classifiers. But the hardware that produces the raw data is at least as important. A new analysis method cannot fix a signal that was altered by a coil change.

Reproducibility crises in psychology and biomedicine have led to reforms in statistical reporting, pre-registration, and data sharing. The next frontier is the material infrastructure. If a field cannot trust that a result obtained on one machine will hold on another, then its claims about the brain are provisional in a way that is rarely acknowledged. The 13-of-20 result is not a reason to abandon resting-state fMRI, but it is a reason to treat every scanner upgrade as a natural experiment — and to analyze it as carefully as any study of the brain itself.