AI Algorithms Play a Pivotal Role in Mass Spectrometry Imaging Data Processing—From ‘Describing Images’ to Intelligent Analysis

How large is the dataset for a single mass spectrometry imaging map? For a tissue section measuring 1 centimetre square, scanned at a pixel pitch of 20 micrometres, the raw data typically ranges from several gigabytes to tens of gigabytes. Each pixel contains between several thousand and tens of thousands of m/z channels, and the entire image is essentially a superposition of thousands of molecular distribution maps. Faced with data on this scale, traditional manual channel-by-channel browsing and subjective judgement are no longer sufficient.

Over the past two years, research into the application of artificial intelligence algorithms in mass spectrometry imaging data processing has seen explosive growth. From molecular annotation to tissue segmentation, and from tumour boundary identification to drug metabolism tracking, machine learning and deep learning models are reshaping the analytical paradigm for mass spectrometry imaging data. This trend has also had a reciprocal impact on the hardware side—the design of mass spectrometry imaging sources and ion sources is adapting to the new data quality requirements imposed by AI analysis.

Data Processing Bottlenecks: The Limitations of Traditional Methods

The traditional analysis workflow for mass spectrometry imaging data typically involves: baseline correction → peak alignment → normalisation → extraction of m/z channels of interest → pseudo-colour imaging → manual interpretation. This workflow presents several key bottlenecks.

Difficulties in molecular annotation. The m/z values obtained from mass spectrometry imaging need to be mapped to specific molecules; traditional methods rely on exact mass matching and database searches. However, a single m/z value may correspond to multiple isomers or isobaric variants, which cannot be distinguished based on mass information alone. Researchers often need to acquire additional MS/MS fragmentation data for confirmation, significantly increasing both the workload and the volume of data.

High degree of subjectivity in spatial pattern recognition. Identifying molecular distribution patterns within tissue (such as the boundary between tumour and normal regions) typically relies on the researcher’s visual judgement and experience. Different researchers may produce different regional classifications for the same dataset, resulting in poor reproducibility.

Underutilisation of high-dimensional data. Each pixel contains thousands of m/z channels, but manual analysis typically focuses only on a small number of known target molecules, leaving the vast majority of the data unused. In fact, these ‘non-target’ molecules may harbour as-yet-undiscovered biomarkers or tissue typing information.

A revolution in analysis driven by machine learning

The introduction of AI algorithms is progressively overcoming the aforementioned bottlenecks.

Automatic molecular annotation. Machine learning-based molecular annotation models can comprehensively utilise precise mass, isotope patterns, MS/MS fragmentation features and spatial co-localisation information to automatically match m/z values to candidate molecules. The latest models have achieved an annotation accuracy of over 80 per cent in the field of lipidomics, significantly reducing the workload associated with manual confirmation.

Unsupervised spatial partitioning. Clustering algorithms (such as k-means, hierarchical clustering and NMF) and deep learning models (such as autoencoders and U-Net) can perform unsupervised spatial partitioning of full-spectrum data, automatically identifying regions of molecular interest within tissues. In tumour tissue imaging, AI models can automatically delineate tumour core regions, infiltrating regions and normal tissue regions based on molecular distribution patterns; the partitioning results show a high degree of consistency with pathologists’ H&E staining interpretations.

Biomarker discovery. Differential analysis models (such as OPLS-DA and random forests) can screen through thousands of m/z channels to identify key molecular features that distinguish different tissue regions, thereby discovering potential biomarkers that might be overlooked by traditional targeted analyses.

Multimodal data fusion. The latest research trend involves the fusion and analysis of mass spectrometry imaging data with histological images (H&E and IHC staining) and genomic data. AI models trained on multimodal data can uncover cross-omics correlations that cannot be detected by a single data modality.

New hardware requirements: an AI perspective on data quality

The reliance of AI algorithms on data, in turn, places new technical demands on mass spectrometry imagers and imaging ion sources.

Signal consistency. Machine learning models are highly sensitive to the consistency of training data. If data produced by the same model of imaging source exhibits significant drift across different experiments, the model’s ability to generalise to new data will be significantly reduced. The signal stability of the imaging source is a fundamental prerequisite for AI analysis.

New Source’s MSI DPI electrospray/photochemical ionisation mass spectrometry imaging source incorporates real-time monitoring and feedback control of VUV light flux in its post-photodetection design, ensuring that the Kr lamp’s output power remains stable (fluctuations<5%) throughout="" the="" imaging="" scan.="" this="" engineered="" assurance="" of="" signal="" stability="" provides="" foundation="" for="" data="" quality="" required="" reliable="" training="" subsequent="" ai="" models.="">

Data integrity through full polarity detection. AI models extract information from full-spectrum data; if the imaging ion source exhibits polarity discrimination—detecting only polar molecules whilst omitting non-polar ones—the AI model is effectively trained on ‘incomplete data’, potentially missing key patterns associated with non-polar molecules.

The post-photoionisation technology in the MSI DPI and MSI LDPI imaging sources enables full polarity detection, ensuring that the data for each pixel is complete across the molecular spectrum. Non-polar molecules (cholesterol, triglycerides, steroid compounds) and polar molecules (phospholipids, choline) are detected simultaneously within the same dataset, thereby enabling AI models to utilise more comprehensive molecular information for pattern recognition and biomarker discovery.

Balancing spatial resolution and data density. High spatial resolution implies more pixels and a larger volume of data, but it also means weaker per-pixel signals and higher noise levels. Given the sensitivity of AI models to the signal-to-noise ratio, the imaging source must strike an optimal balance between resolution and sensitivity.

The MSI LDPI imaging source maintains detectable signal levels at a spatial resolution of 2–3 μm through post-electron ionisation technology, enabling imaging data at the single-cell scale to possess both spatial accuracy and a signal-to-noise ratio that can be effectively utilised by AI models. This quantified, high-resolution data is precisely the input required to train cell-level spatialomics AI models.

From Data to Diagnosis: The Clinical Pathway of AI + Mass Spectrometry Imaging

The combination of AI algorithms and mass spectrometry imaging is paving the way for translation from the laboratory to the clinic.

Intraoperative margin assessment is one of the scenarios closest to clinical practice. During surgical resection of tumours, the presence or absence of residual tumour cells at the resection margin directly affects the prognosis. Traditional intraoperative frozen section pathology assessment takes 30–45 minutes and relies on the pathologist’s expert judgement. The theoretical workflow for mass spectrometry imaging combined with AI models is as follows: surgical resection of tissue → rapid scanning via mass spectrometry imaging → real-time interpretation of molecular characteristics at the resection margin by the AI model → intraoperative determination of the tumour/normal boundary. If the total duration of this workflow can be reduced to within 15–20 minutes, it will possess practical value for intraoperative auxiliary diagnosis.

Implementing this workflow requires the synergy of hardware and software. On the hardware side, a high-speed mass spectrometry imager and a high-throughput imaging ion source are required—Xinceyuan’s MSI DPI imaging source optimises the response speed of the atomisation–transport–detection chain, reducing the dwell time per pixel to 0.1–0.2 seconds, thereby providing the hardware foundation for shortening the total imaging time. On the software side, clinically validated AI models are required—this relies on the accumulation of large amounts of annotated data (tissue imaging data with known tumour/normal boundaries) and model training.

Industry Outlook

The trend towards the integration of AI algorithms into mass spectrometry imaging data processing is now irreversible. For manufacturers of mass spectrometry imagers and imaging ion sources, this trend offers two strategic insights:

Firstly, the evaluation criteria for hardware performance are expanding from ‘single-detection capability’ to ‘how well the data quality supports AI models’. Characteristics such as signal stability, full polarity coverage and the balance between spatial resolution and sensitivity will directly influence the effectiveness of downstream AI analysis, becoming a new dimension of product competitiveness.

Second, collaboration between hardware manufacturers and AI algorithm teams will become a competitive advantage. If imaging source manufacturers can establish data-sharing and joint optimisation mechanisms with AI model development teams—enabling hardware designs to be specifically optimised for AI analysis requirements—they will gain a head start in the industrial convergence of “AI + mass spectrometry imaging”.

The future of mass spectrometry imaging technology depends not only on ‘how clearly we can see’, but also on ‘how deeply we can understand’. AI algorithms are becoming the core engine driving the latter.