Unlocking High-Dimensional Insights in LA-ICP-TOF-MS Imaging

Elemental mass spectrometry imaging (MSI), particularly laser ablation-inductively coupled plasma-time-of-flight mass spectrometry (LA-ICP-TOF-MS), produces highly complex, pixel-resolved datasets containing full mass spectra across hundreds of m/z values. While quadrupole instruments once dominated, their sequential scanning limits isotopic coverage. TOF technology overcomes this by capturing complete spectra per pixel at kilohertz rates and micrometer resolution, generating information-rich, high-dimensional data. Manual examination of individual m/z images is no longer practical, especially for non-targeted studies where key spatial or elemental patterns may be overlooked. Each pixel functions as a point in multidimensional space and modern dimensionality-reduction algorithms now offer powerful solutions for visualizing, segmenting, and interpreting these complex datasets. Algorithms such as uniform manifold approximation and projection (UMAP), which are well-suited for preserving both local structure and global trends, enable efficient visualization, segmentation, and interpretation of these intricate datasets, providing a powerful pathway to insights not accessible through manual analysis.

A recent study published in the Journal of Analytical Atomic Spectrometry (1) provides a detailed introduction to UMAP for analyzing LA-ICP-TOF-MS data. By converting high-dimensional MSI data into two-dimensional spaces, UMAP facilitates automated visualization to identify spectral clusters. Spectroscopy spoke to the paper’s lead author, Katharina Kronenberg of the University of Graz, about her group’s work.

What distinguishes laser ablation-inductively coupled plasma-time of flight mass spectrometry (LA-ICP-MS) from traditional solution-based ICP-MS in terms of sample introduction and data structure?

The laser ablation system facilitates the introduction of solid samples into the ICP-MS. The specimen is situated within an ablation cell, which is sustained under a helium (He) atmosphere. When the laser fires, an aerosol of particles is generated that is transported via a He flow to the ICP-MS. While solution-based ICP-MS is inherently operated under wet plasma conditions, LA is most used with a dry plasma. In both LA-ICP-MS and solution-based ICP-MS, data is acquired in a time-resolved manner. However, when LA is employed in a mapping mode wherein the ablation cell with the sample is in motion while the laser is active, the acquired time-resolved data is converted into spatial information. This approach facilitates the generation of elemental maps, which provide crucial insights into the spatial distribution of the elements of interest.

How do quadrupole and TOF mass analyzers differ in data acquisition, and why is TOF particularly advantageous for imaging applications?

In quadrupole-based ICP-MS, the quadrupole acts as a mass filter that sequentially scans the mass-to-charge ratios (m/z) of interest. Consequently, the detection of ions is inherently sequential. As a result, there is a trade-off between the number of isotopes measured and the temporal resolution of the instrument. In LA mapping applications, the signal intensity can fluctuate rapidly, due to spatially heterogeneous samples and fast ablating LA systems. For capturing these transient signals, the sequential nature of quadrupole scanning makes it less suitable. Consequently, in quadrupole-based LA-ICP-MS, the number of isotopes that can be monitored is limited to only a few m/z to maintain adequate temporal resolution.

In contrast, TOF mass analyzers allow the rapid and quasi-simultaneous acquisition of a full mass spectrum per pixel. Within the TOF mass analyzer, the ion beam is cut into ion packets that are accelerated so that all ions attain the same kinetic energy, causing their velocities to differ according to their m/z. A mass spectrum is obtained by recording the TOF and intensity of ions from the entire mass range for each ion packet. This fast detection capability makes TOF-MS particularly advantageous for imaging applications, where high-speed, multi-elemental analysis is essential to accurately capture the fast transient signals and spatial variations in the sample.

Why is dwell time an important parameter in LA-ICP-MS, and how does it influence the quality and accuracy of elemental images?

Dwell time is a key parameter in quadrupole-based LA-ICP-MS because it controls how long the MS measures each m/z before moving to the next. All selected m/z values are measured sequentially, and the sum of all dwell times forms the duty cycle. Each laser shot generates a peak that must be measured in a time-resolved manner. The faster the washout of the plume of particles generated by the laser shot, the shorter and narrower is the peak that must be time-resolved by the MS. Consequently, for fast ablation, fast washout is needed and less time can be spent per peak and per duty cycle. As a result, for quadrupole ICP-MS, also the dwell time per m/z must be shorter, resulting in a worse signal-to-noise ratio and image quality. Alternatively, fewer m/z values can be selected per duty cycle to maintain a suitable signal-to-noise ratio, but this comes at the cost of less elemental information. A TOF-based ICP-MS system, however, is fast in acquiring time-resolved data for the entire mass spectrum and therefore is better suited for fast multi-elemental imaging.

Describe the main challenges associated with non-targeted analysis in LA-ICP-TOF-MS datasets.

In non-targeted analysis, often no a priori knowledge of the sample exits. When performing non-targeted analysis with LA-ICP-TOF-MS, the entire mass range is selected to capture all elements that may potentially be present in the sample. For example, the mass range can be set from ⁷Li to ²³⁸U, resulting in full mass spectra with a large number of recorded m/z. For each recorded m/z in LA-ICP-TOF-MS datasets, an image is generated. The manual and visual examination of this substantial number of elemental images becomes time-consuming and impractical. Additionally, increasing dataset complexity greatly amplifies the risk of overlooking relevant spatial and spectral features. In particular, it becomes challenging to segment spatial regions with distinct multi-elemental signatures or to visualize and interpret overarching patterns within the data.

How does dimensionality reduction, such as UMAP, help in interpreting complex high-dimensional MSI datasets?

To answer this question, it is important to understand the structure of MSI data. In MSI, the intensity of each recorded m/z value represents a variable and therefore a distinct dimension. For example, in a dataset of 200 m/z values, each pixel contains an intensity value for each of the 200 m/z values and is positioned accordingly in a 200-dimensional scatterplot. This scatterplot represents the image in a high-dimensional space.It is difficult to visualize and interpret a dataset with many dimensions directly, as visualizing more than a few dimensions simultaneously is inherently limited. Dimensionality reduction algorithms, such as UMAP, aim to simplify this high-dimensional data structure by projecting it into a low-dimensional (2D or 3D) space,while preserving the underlying relationships within the data structure. This low-dimensional representation of the data, also known as embedding, becomes more interpretable to the human eye. Consequently, dimensionality reduction supports an automated and data-driven approach for identifying clusters with spectral similarity and enabling segmentation of the image based on these clusters.

In LA-ICP-TOF-MS imaging, each pixel can be represented in a 200-dimensional space. How would you approach analyzing or clustering such data to identify regions of interest?

In general, UMAP can be used to identify regions of interest in either a targeted or a non-targeted way. When applied in a non-targeted manner, the user would incorporate the entire dataset for the UMAP run, including all pixels and recorded m/z values. The embedding, which is the low-dimensional output of a UMAP run, will represent the data structure by spectrally similar pixels being packed together. When defining groups in this embedding, these pixels can be mapped back on to the image revealing the spatial position of these spectrally similar pixels. This presents an interactive and exploratory way to recognize patterns and divide the sample into regions of interest. Additionally, regions of interest can be identified in a targeted manner. For instance, when the desired region of interest is a fine spatial structure, such as dispersed nanoparticles, perforated tissue or geological inclusions. Manually drawing these regions in an image would be tedious and possibly inaccurate. However, one may already know the elemental content of these structures. This knowledge can be leveraged in UMAP runs by restricting the input to the relevant m/z and/or pixel region where these structures appear. By iteratively adjusting parameters and input features, UMAP can be guided to form clusters that correspond to desired spatial regions of interest.

What roles do the UMAP parameters n_neighbors and min_dist play in controlling the balance between local and global structure in the embedding?

In the UMAP algorithm, n_neighbors and min_dist are the two most important hyper-parameters because they control how the algorithm balances local and global data structure while the data is projected into the low-dimensional embedding. The parameter n_neighbors influences the first stage of the algorithm, where UMAP builds a graph that represents the topological structure of the data in high-dimensional space. It determines how many nearest neighbors are considered for each point, and how much local versus global structure is preserved. Smaller values (such as 5–15) emphasize fine-grained, local patterns, while larger values (such as 50–200) help maintain more global, large-scale groupings. If n_neighbors is too small, UMAP may capture noise. If it is too large, it may over-smooth the data and miss important local patterns. Because it depends on the distribution of the data, it must be re-evaluated for each new dataset.

The parameter min_dist affects the second stage of the algorithm, where UMAP optimizes the low-dimensional embedding. It sets the minimum distance between points in the embedding and therefore controls how densely points within groupings are arranged. Low values create tightly packed clusters, while larger values spread points further apart, influencing the visual density and appearance of the final embedding.

Can you explain the concept and benefit of hierarchical UMAP in segmenting fine spatial structures within MSI datasets?

Hierarchical UMAP is a workflow in which an initial UMAP embedding is used to identify major patterns, and individual clusters of interest are then isolated and re-analyzed with UMAP to reveal finer spectral substructures. After the first UMAP run, the dominant spectral differences within the sample are identified and divided into different clusters. When a cluster is considered to contain further unresolved internal structure, this region can be extracted and processed again with UMAP. This second embedding can uncover distinct subclusters that might not be separable in the initial embedding of the entire dataset. Hierarchical UMAP analysis can also be conducted in either a targeted or non-targeted manner.

Describe how UMAP-based analysis can reveal anatomical or tissue-specific elemental patterns that may not be evident from individual isotope images.

UMAP-based analysis can reveal anatomical or tissue-specific elemental patterns by integrating the complete multi-elemental information from every pixel, rather than relying on what can be visually inferred from single-isotope images. When a single elemental map is inspected in isolation, regions with very strong signals tend to dominate the color scale, often causing more subtle but biologically meaningful variations elsewhere in the sample to be overlooked. Even when an element appears similarly abundant in two spatial regions, their overall elemental profiles may differ substantially. These differences are not apparent from the individual isotope image alone. Because UMAP embeds pixels on the basis of their full elemental profile, it can separate tissue regions that look similar in one isotopic map but differ across the mass spectrum. At the same time, it highlights the presence of widespread but low-intensity elemental distributions that might remain invisible when color scales are automatically adjusted to the most intense region. By simultaneously considering all intensity values from all recorded isotopes, UMAP reduces the bias introduced by visual interpretation and color-scale choices, and thus uncovers tissue-specific elemental patterns that would not be evident from single-isotope images inspected by eye.

If given LA-ICP-TOF-MS imaging data from an unknown biological tissue, outline your analytical workflow from raw data preprocessing to visual interpretation using dimensionality reduction methods.

In a first step, I import the raw dataset into the newly developed Multiscale-Image-Analysis (MIA) software, which subtracts the gas-blank background for each ablated line. The user interface comprises an image viewer, a spectrum viewer, and an embedding viewer.

Next, I initiate a non-targeted UMAP run that includes all recorded m/zvalues and all pixels of the dataset. To achieve tightly packed groupings in the embedding, min_dist should be set to a small value (for example, 0.1). The parameter n_neighbors needs to be empirically optimized for each dataset, but a reasonable starting range is between 5 and 20. The resulting embedding is then displayed in the embedding viewer.

In this embedding, each pixel is represented as a single point. If the sample contains spectrally distinct regions, pixels with similar spectral profiles will form a cluster. Using an interactive drawing tool, groupings of points can be segmented and assigned different colors. At the same time in the image viewer, the pixels of each segment adopt the corresponding colors. Additionally, the spectrum viewer displays the mean mass spectra for the different segments with their respective colors, enabling assessment of spectral differences between regions. Dominant m/zvalues can also be selected for image display to visualize their spatial distribution within the sample.

Following this, a second and more refined UMAP run can be performed. For example, based on the initial UMAP result, I would seek to distinguish between the tissue region and the background of the tissue-section substrate. The second run could therefore include only pixels originating from biological tissue. Furthermore, the spectrum viewer highlights dominant m/z values, which can be used to further refine the second UMAP analysis. A list of important m/zvalues can be generated by clicking in the spectrum, and this list serves as the basis for the second UMAP run. I employ such refinement strategies to enhance the differentiation of spectrally distinct regions in my analysis workflow.

Depending on the analytical objective, a hierarchical UMAP approach may subsequently be applied to explore the sample in even greater detail. This workflow allows for interactive and data-driven segmentation and exploration of the MSI data. As with any data analysis workflow, it is important to assess whether the segmentation results align with existing knowledge regarding the sample’s structure, for example by comparing them to a microscopic image.

To interactively explore LA-ICP-TOF-MS datasets with UMAP, we developed a new software (Hennes Rave, MIA: Multiscale Image Analysis, 2025): https://github.com/hennesrave/multiscale-image-analysis).

References

1. Kronenberg, K.; Rave, H.; Ghaffari-Tabrizi-Wizsy, N. et al. Exploring High-Dimensional LA-ICP-TOFMS Data with Uniform Manifold Approximation and Projection (UMAP).J. Anal. At. Spectrom. 2025, 40, 3473-3484. DOI: 10.1039/D5JA00215J