Ultra-Low-Field MRI Can Support Brain Network Mapping Without Losing Core Signal
Ultra-low-field MRI is usually discussed in terms of access: smaller systems, easier deployment, and the possibility of point-of-care neuroimaging outside the traditional high-field scanner environment. This preprint pushes the question further. Can a 64 mT system support the multivariate morphometric network analyses that have become central to modern structural connectomics? The answer, in this first feasibility and reliability assessment, is yes — with important caveats about which analytic pipeline is being used and what the investigator wants the network to do.
Two network families formed the core of the evaluation: Morphometric Similarity Networks (MSNs) and Morphometric Inverse Divergence (MIND) networks. Both were derived from paired high-field and ultra-low-field scans in a primary cohort of 23 participants, then examined again in an independent replication cohort of 19 participants. The study tested whether the networks could be reconstructed consistently across field strengths, and whether processing choices such as super-resolution, cortical parcellation granularity, and mesh resolution altered stability or identifiability. In other words, the project did not just ask whether the scans looked usable. It asked whether the resulting graph structure survived the jump from conventional MRI to an ultra-low-field platform.
MIND Networks Held Up Best Across Field Strengths
The clearest signal came from the edge-level agreement data. MIND networks performed better than MSNs when high-field and ultra-low-field data were compared directly. Mean concordance for MIND reached 0.95 across scanner types, with ultra-low-field reproducibility at 0.98. MSNs were also workable, but their agreement was lower, with mean concordance in the 0.86 to 0.87 range and reproducibility between 0.91 and 0.94.
That gap matters because edge-level concordance is the basic test of whether a network representation preserves the same topological relationships when the acquisition environment changes. Ultra-low-field systems operate under different signal constraints, so the fact that MIND retained near-ceiling reproducibility suggests the method is comparatively robust to the field-strength divide. In practical terms, the morphometric signature generated by MIND appears less vulnerable to the noise and resolution penalties that usually accompany lower-field imaging.
The study’s design also makes the result harder to dismiss as a single-cohort artifact. The replication cohort was independent, which strengthens the case that the observed stability is not just a feature of one sample or one processing path. For an imaging method that aims to become globally scalable, that matters as much as the raw coefficient itself.
MSN Favored Individual Uniqueness Over Cross-Scanner Stability
The result was not one-dimensional. While MIND delivered stronger cross-scanner agreement, MSNs produced higher subject-level differential identifiability across field strengths. The reason was conceptual and biologically intuitive: MSN generated a broader separation between within-subject and between-subject similarity, which makes individual network fingerprints easier to distinguish.
This is where the bedside and research uses begin to diverge. A method optimized for scanner-to-scanner stability may be the better choice for multicenter standardization, while a method that accentuates individual uniqueness may be more useful for patient stratification, longitudinal phenotyping, or subject-level biomarker development. Structural connectomics often faces a trade-off between reproducibility and distinctiveness. Here, the data show that ultra-low-field MRI can support both goals, but not with the same analytic emphasis.
The finding also fits with how morphometric network methods work biologically. MSNs rely on interregional similarity patterns that can preserve personal anatomical features, while MIND transforms those relationships in a way that appears to smooth over some of the scanner-dependent variability. The two approaches are therefore not redundant. They are different readouts of the same underlying anatomy, and they respond differently to the constraints of low-field acquisition.
Processing Choices Changed Performance in Predictable Ways
The study did not stop at method comparison. It also examined how reconstruction choices changed performance. Rebuilding MIND networks on native high-resolution meshes, or fsnative, maximized cross-scanner agreement. Coarser meshes and finer cortical parcellations, by contrast, improved individual network identifiability.
This is a useful reminder that the best pipeline depends on the scientific question. Preserving native mesh detail appears to help MIND retain stability across field strengths, which makes sense if the goal is to minimize interpolation-related distortion. Finer parcellation can sharpen the separability of individual patterns, because more granular regional definitions may capture subject-specific morphometric variation that is diluted when regions are merged. The mesh and atlas choices are not technical footnotes; they shape whether the network becomes more reproducible or more individual-specific.
The work therefore points toward a split use case for ultra-low-field connectomics. If investigators want cross-site comparability, MIND on native meshes is the safer path. If they want stronger subject-level discrimination, MSN with finer parcellation may be preferable. The platform itself is not the limiting factor. The analytic objective is.
Why This Feasibility Study Matters for Access and Neuroimaging Design
Ultra-low-field MRI carries obvious logistical appeal. It may be easier to deploy in settings where conventional scanners are unavailable, impractical, or too costly to scale. The clinical and research interest in such systems has mostly centered on basic image acquisition and gross anatomy. This preprint extends the discussion into a more sophisticated domain: whether a low-field scanner can support morphometric graph analysis, a technique that depends on subtle regional relationships rather than simple visual interpretation.
If that capability holds up in larger studies, the implications are broad. Connectomic research could become less dependent on a narrow set of high-resource imaging centers. Longitudinal studies might be able to recruit and follow participants in more accessible environments. Neurodegeneration research, in particular, could benefit from a modality that makes structural network mapping more practical outside tertiary hospitals.
For now, the evidence is strongest at the level of feasibility and reliability, not clinical validation. The datasets were modest in size, and the work is still a preprint rather than a peer-reviewed paper. The study also focuses on technical performance metrics — concordance and identifiability — rather than disease prediction, symptom correlation, or outcome tracking. Those next steps will determine whether ultra-low-field morphometric networks remain an imaging curiosity or become a usable tool in larger neuroimaging programs.
The most important point is narrower than a headline about accessibility would suggest: ultra-low-field MRI can preserve meaningful morphometric network structure, but the choice between MIND and MSN changes what kind of structure is preserved best. One approach favors stability across scanners. The other favors individual distinctiveness. The field now has a concrete starting point for deciding which property matters most in the next generation of scalable brain network research.