MSACL 2026 Abstract

A Workflow for Multimodal Analysis of Mass Spectrometry Imaging and Pathology Foundation-Model Feature Maps

Amoon Jamzad, Martin Kaufmann, John F. Rudan, Sonal Varma, Gabor Fichtinger, Parvin Mousavi
Queen’s University, Kingston, ON, Canada

Amoon Jamzad, PhD (Presenter) Queen's University

Presenter Bio: Dr. Amoon Jamzad is an AI research scientist and engineer focused on machine learning for biomedical and clinical data. His work centers on translating complex data, particularly mass spectrometry imaging, into clinically useful insights through computational methods and practical software tools. He is currently a Postdoctoral Fellow and Adjunct Assistant Professor at School of Computing, Queen’s University. He is the developer of MassVision, an open-source platform for MSI analysis, and his work has appeared in venues including Analytical Chemistry, MICCAI, and IJCARS.

Relevant Financial Disclosures (within past 24 months, reported on Apr 22, 2026) No relevant financial relationship(s) to disclose.

INTRODUCTION:
Mass spectrometry imaging (MSI) and whole-slide pathology imaging (WSI) provide complementary views of tissue, with MSI capturing spatial molecular composition and WSI capturing morphology. Recent pathology foundation models can encode WSI patches into informative feature representations, creating new opportunities for AI-driven tissue analysis. However, multimodal MSI studies remain challenging because pathology-derived features and MSI are not naturally represented in a common spatial framework that supports direct comparison and AI-driven analysis.

OBJECTIVE:
To develop an end-to-end workflow that transforms pathology foundation-model outputs into spatial feature maps that can be analyzed together with MSI for multimodal visualization, spatial colocalization, paired dataset generation, and AI-based analysis.

METHODS:
Previously reported DESI MSI data from 10 colorectal tissue slides [1,2] were paired with corresponding WSI data. WSI patches were extracted at 100 µm to match the DESI pixel size and encoded with Virchow2 [3] pathology foundation model to generate patch-level embeddings. Rather than treating these embeddings as isolated patch descriptors, they were reorganized by spatial location to create pathology feature maps analogous to MSI molecular data cubes. This step was central to the workflow because it enabled pathology-derived AI features to be explored in the same spatial format as MSI. The pathology feature maps and MSI data were imported into MassVision [1], an open-source platform for interpretation and analysis of high-dimensional spatial data. Modality-specific visualizations were generated and used for landmark-based registration. Regions of interest were then annotated in one modality, and the resulting transformation was applied to extract matched MSI and pathology-embedding datasets from corresponding tissue locations. This process yielded 16,677 matched spectrum-patch pairs across six tissue classes: adenocarcinoma, benign mucosa, serosa, smooth muscle, submucosa, and inflammatory cells. Preliminary classification was performed in MassVision using PCA-LDA with leave-one-patient-out cross-validation. In addition, to assess the effect of morphologic context, pathology features were extracted from larger concentric patches of 300 and 500 µm centered on the original 100 µm patches.

RESULTS:
The resulting paired molecular-morphologic representation supported quantitative tissue discrimination in both modalities and enabled direct comparison of their predictive behavior. The DESI-based tissue classification achieved a mean accuracy of 83.97 ± 7.65% in the six-class task. Classification using pathology feature maps showed comparable performance at the matched 100 µm scale (83.02 ± 4.19%) and improved as larger contextual fields were incorporated, reaching 89.04 ± 4.08% at 300 µm and 91.16 ± 4.25% at 500 µm. Together, these results show that representing pathology embeddings as spatial feature maps creates a practical bridge between molecular and morphologic tissue analysis, while the gains observed at larger pathology scales highlight the value of contextual tissue information for multimodal MSI studies.

CONCLUSIONS:
This work presents a practical workflow for multimodal MSI studies in which digital pathology is transformed from a static reference image into an AI-derived spatial companion modality. By representing pathology foundation-model embeddings as feature maps, the workflow enables joint visualization, spatial colocalization, pathology-guided label transfer, and paired molecular-morphologic analysis. Demonstrated here on colorectal DESI data, this approach establishes a foundation for future multimodal modeling and morphology-guided MSI learning.

REFERENCES:
[1] Jamzad A, et al. MassVision: An Open-Source End-to-End Platform for AI-Driven Mass Spectrometry Imaging Analysis. Analytical Chemistry. 2025;97(39):21588-21597.
[2] Kaufmann M, et al. Metabolically Active Zones Involving Fatty Acid Elongation Delineated by DESI-MSI Correlate with Pathological and Prognostic Features of Colorectal Cancer. Metabolites. 2023;13(4):508.
[3] Zimmermann E, et al. Virchow2: Scaling Self-Supervised Mixed Magnification Models in Pathology. arXiv. 2024:2408.00738.