Performing MS imaging histology coregistration well is what converts a grid of mass spectra into something a biologist can actually read — an ion image alone shows where a mass-to-charge value was detected, not what tissue structure that corresponds to. Overlaying it precisely onto a stained section, or onto another spatial modality entirely, is where interpretation becomes possible.
Key Takeaways
|
Why Coregistration Matters
An MS image is a grid of spectra with coordinates — and nothing in that grid says which coordinates correspond to epithelium, stroma, or necrosis. Histological staining is what supplies that context, but staining is generally performed on a separate, consecutive section rather than the one that was ionised, because the preparation each requires differs, and because the imaged section has often been physically altered by the acquisition itself.
That separation creates the registration problem. Two images of what is nominally the same tissue, acquired under different conditions and sometimes from adjacent rather than identical material, have to be brought into a shared coordinate system before a reader can ask whether a given molecular signal sits within a specific histological structure. Getting this wrong produces a plausible-looking but incorrect answer, which is a more dangerous failure mode than an obviously broken image.
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
One complication is worth flagging before any registration technique: sample preparation itself can move analytes. As covered in MALDI Imaging Mass Spectrometry: How It Works, matrix application and, for DESI, the sprayed solvent, can cause delocalisation, so the choice of which ion image to register, and an awareness that its boundaries may not perfectly reflect the underlying tissue, has to be made with that possibility in mind rather than assumed away.
Registering MSI to H&E
The classical approach uses landmarks, also called fiducial markers or control points: features visible in both images, manually identified and placed by an operator, from which a geometric transformation between the two coordinate systems is calculated.
That approach has an elegant variant worth knowing about specifically.
The Instrument Writes Its Own Landmarks Rather than asking an operator to find matching features in two dissimilar-looking images, one method exploits something MALDI produces as a side effect of running: laser ablation leaves visible marks on the tissue surface, and those marks appear in both the MS acquisition and a subsequent optical image of the same section. Using them as ad-hoc landmarks means the registration reference points already exist in both modalities without any additional labelling step. The published methodology for validating this approach is worth adopting as a general practice, not only for this specific technique: co-registration error was estimated by counting the number of pixels in the optical image, at approximately 2 micrometres per pixel, between the centre of a laser-shot landmark and the corresponding MSI pixel. Stating accuracy in physical units against a defined reference, rather than asserting that registration was successful, is what turns a workflow into a validated method. |
From those landmarks, or from any set of control points, a geometric transformation is calculated and applied. Two families of transformation are available, and the choice matters.
Transformation | What It Assumes | When It Is Adequate |
Rigid or affine | Tissue moved as a single unit: some combination of translation, rotation, scaling, and shear applied uniformly across the whole image | Minimal local distortion, and when only a single ion image is being aligned to a roughly similarly shaped optical image |
Non-rigid | Different regions of the tissue may have deformed differently, for example through local tearing, stretching, or folding during processing | Any situation where local deformation is likely, which increases with section size, tissue fragility, and the time elapsed between the two acquisitions |
Table 1. The two families of registration transformation. Most limitations attributed to older tools trace to being restricted to the top row when the tissue actually needed the bottom one.
Landmark-based rigid or affine registration is also where an operator dependency creeps back in, since manual placement introduces variability and does not scale well to large studies. A landmark-free R package developed specifically for MALDI-to-H&E registration addresses both limitations at once: it requires no manually placed landmarks at all, works directly from image content, is not tied to any particular ionisation source, and extends to non-rigid transformations that account for the local deformation that purely affine methods miss. That combination, no manual step and deformation awareness, is the direction registration methods have moved in.
Multimodal Overlays: MSI Plus Other Modalities
Combining MS imaging with another spatially resolved technique, such as immunofluorescence or spatial transcriptomics, multiplies the value of both but forces a decision that pure MSI-to-H&E registration does not: whether to run both modalities on the same section, destructively, or on adjacent serial sections, non-destructively but with an added registration burden between sections that were never identical to begin with.
One described same-section approach uses the physical evidence MSI leaves behind directly: a MATLAB-based tool imports cell boundary coordinates from an imaging-based spatial transcriptomics platform and uses the ablation micro-dots from the MS run as internal fiducial markers, achieving pixel-scale coregistration, then extracts a per-cell mass spectrum by assigning each MSI pixel to its overlapping cell boundary polygon. That directly links a metabolic readout to a transcriptomic identity for the same physical cell, which is the strongest possible form of multimodal linkage, because there is no intersection registration error to account for at all.
What Same-Section Analysis Costs the Other Modality The obvious concern with running MS imaging and transcriptomics on the same section is that laser ablation is destructive, and destroying tissue before extracting RNA from it should degrade the sequencing result. One described implementation of this approach reports a specific and reassuring answer: it introduces approximately 20 to 30 percent fewer transcript counts per cell due to ablation, but overall transcriptome complexity and cell-type classification accuracy remain preserved. |
Where same-section analysis is not viable, because the two modalities have genuinely incompatible preparation requirements, serial-section registration is the fallback. A described workflow for this situation registers MSI and spatial transcriptomics data acquired from adjacent sections using a non-rigid algorithm with automated background segmentation and centre-of-mass preprocessing to align stained histology anchors between the two. That approach carries an additional layer of registration uncertainty, since the sections are not identical, but it avoids the destructive trade-off entirely.
Downstream of registration, whichever route produced it, the analytical work of cell typing, deconvolution, and pattern detection across the combined dataset draws on methods and toolkits that are not specific to mass spectrometry. Our sister publication, Technology Network's Guide to Analyzing Spatial Biology Data, covers that broader computational layer, including cell segmentation and typing workflows for sequencing- and imaging-based platforms, which is genuinely useful context once a multimodal dataset has been registered and is ready for joint analysis.
Common Pitfalls
Five failure modes recur, and most are avoidable with a specific check rather than general care.
- Registering a delocalised ion image without checking for it. If matrix application or solvent-based ionisation moved an analyte before acquisition, the boundary of that ion signal does not represent the true tissue boundary, and registering it precisely to histology only makes the error look authoritative.
- Forcing an affine transform on a deformed section. Larger sections, fragile tissue, and longer gaps between the two acquisitions all increase the chance of local deformation that a rigid transformation cannot represent, producing systematic local misalignment that a global accuracy metric can hide.
- Reporting that registration succeeded without a quantified accuracy check. A visual impression of alignment is not evidence. Counting the pixel distance between known corresponding points, as described above, is a concrete and repeatable check that costs little.
- Assuming same-section multimodal analysis is free. Ablation is destructive, and even where the biological result survives, the raw signal is reduced. Plan for that reduction rather than discovering it in the data.
- Treating serial sections as identical. Adjacent sections differ by the section thickness at minimum, and often by more due to local tissue heterogeneity. Registration error between serial sections should be estimated, not assumed negligible, particularly for small structures.
For where coregistration sits within the overall preparation and analysis sequence, see Sample Preparation for Spatial Analysis: From Tissue to Data. The wider question of how much data a registered multimodal dataset actually represents, and the arithmetic behind acquisition time at the resolution such registration demands, is developed in Spatial Resolution vs. Sensitivity in MS Imaging: The Fundamental Trade-off. Detailed treatment of the downstream analysis, software, and machine learning approaches that follow registration is in Analyzing Mass Spectrometry Imaging Data: Processing, Statistics, and Multimodal Integration.
This article was produced under Separation Science's AI Editorial Guidelines.
