When Mass Spectrometry Imaging Meets Artificial Intelligence: From Point-by-Point Scanning to Automated Image Interpretation

Mass spectrometry imaging is reaching a turning point: over the past decade, the industry has primarily focused on addressing the question of ‘whether imaging is possible’; the focus will now gradually shift to ‘how to interpret the resulting images’. And it is artificial intelligence that is capable of handling this vast volume of image data.

The challenge lies not in acquisition, but in interpretation

A single high-resolution imaging run produces three-dimensional data comprising ‘spatial pixels × m/z channels’, often amounting to gigabytes. In traditional workflows, researchers must manually select m/z values of interest, compare images one by one, and rely on experience to determine which areas appear ‘different’. Whilst this method is manageable with a small number of samples, it becomes a bottleneck as the sample size increases.

This is precisely where artificial intelligence comes into play—it does not seek to replace the ionisation source, but rather takes over the latter half of the process: ‘image analysis’.

Three emerging directions

Automatic difference detection. By applying unsupervised learning methods to perform dimensionality reduction and clustering on the entire imaging dataset, the system can automatically flag ‘regions with markedly different molecular compositions’ without relying on manual guesswork regarding m/z values. This shifts the approach from ‘hypothesis-driven’ to ‘data-driven’.

Cross-slide registration. Imaging of the same tissue across different slides or batches often exhibits spatial coordinate drift. Image registration algorithms can align these images, enabling longitudinal comparisons and stabilising multi-omics fusion.

Modal fusion. By training models using molecular information from mass spectrometry imaging alongside optical images and histopathological annotations, the models gradually learn to “associate molecular patterns with histological significance”. This has direct scientific value for tumour classification, drug efficacy assessment and other applications.

Unavoidable practical constraints

Firstly, data standards. Spectral resolutions and coordinate systems vary across different imaging sources and platforms; without unified pre-processing specifications, it is difficult for models to generalise across devices. Secondly, interpretability. Both research and clinical settings often require an explanation of ‘why this judgement was made’; purely ‘black-box’ conclusions are unlikely to be accepted. Thirdly, computing power and storage. Reconstructing and inferring from gigabyte-scale datasets within a reasonable timeframe places demands on local computing power.

Neo Source INSTRUMENT’s imaging sources provide resolution (DPI 20–200 μm, LDPI 2–3 μm) and compatibility with Agilent, AB SCIEX and Thermo platforms as core capabilities; this effectively delivers ‘clean, reproducible’ imaging data to users—which is precisely the prerequisite for subsequent integration with AI workflows.

Outlook

In the short term, AI will primarily serve as an ‘auxiliary image reader’ for researchers, handling registration, dimensionality reduction and initial screening; in the medium term, automatic annotation and cross-modal correlation will become part of standard workflows; in the longer term, imaging, analysis and interpretation may be consolidated into a near-real-time workflow.

However, no matter how intelligent the technology becomes, ‘whether the ionisation is stable, whether the positioning is accurate, and whether the platform is compatible’ remain the bedrock. AI will not render imaging sources redundant; rather, it will enhance the value of a reliable, reproducible imaging system—because good models always require good data to be trained on.