Manifold Learning for Visualization, Prioritization, and Data Fusion of Mass Spectrometry Imaging Data

Mass Spectrometry Imaging (MSI) is a powerful molecular imaging technology that can detect the spatial distribution of molecules in a tissue section. Because MSI does not require any a priori labeling, the technique has become very popular for the explorative comparison of metabolites, lipids, peptides and proteins between various tissue regions. Since it has been shown that tumor heterogeneity plays an important role in tumor biology, it has become clear that we need to elucidate the spatial distribution of molecules. MSI can therefore be of significant relevance in predicting cancer progression and treatment response, which often remains a challenge in today's clinical practice. A single measurement can however lead to complex and high dimensional data with file sizes in the gigabyte and even the terabyte range. As such manually exploring the data is becoming infeasible and support from computational methods is required. The focus of this work is therefore the development and application of computational methods to MSI data. Specifically we have concentrated on the topics related to non-linear dimensionality reduction, the prioritization of molecules measured per tissue region, and data fusion of MSI with the corresponding histology or microscopy image. For the non-linear dimensionality reduction, we make use of Uniform Manifold Approximation and Projection (UMAP) through which we achieve excellent visualizations of MSI data as demonstrated by the corresponding histology or microscopy images. We have conducted an extensive evaluation regarding the performance and results in comparison to other dimensionality reduction methods. To this end, we have used spatial autocorrelation and spectral similarity as a benchmark. In addition, we have empirically evaluated a number of different distance metrics, where we show that the choice of a particular distance metric might impact the visualization outcome. Building further on the obtained visualizations using UMAP, we proposed a bi-directional dimensionality reduction approach to prioritize the molecules driving these observations. This approach enables the prioritization of m/z-values in individual tissue samples but also across different tissue samples through the incorporation of both spatial and spectral information. The approach was demonstrated for tissue samples obtained from healthy mouse pancreas tissue. Finally, we introduce the correspondence-aware manifold learning paradigm for data fusion of molecular imaging data with the corresponding microscopy images. This enables us to bring the molecular information to a higher spatial resolution. As such these visualizations can play an important role in the digital pathology field for the quick assessment of a complete MSI dataset from a pathologist's perspective. We have shown that using this approach it becomes possible to identify a single infiltrating plasma cell amongst a group of phenotypically different (epithelial) cells. The identification of single aberrant cells is of crucial importance to evaluate a wide range of pathologies, in particular in cancer.

Jaar van publicatie:2021

Toegankelijkheid:Closed