Exploring the Efficacy of Background Removal with SOM-RPM in Mass Spectrometry Imaging

Image segmentation is a critical yet fundamental aspect of interpreting the spatio-spectral information contained in complex, high-dimensional mass spectrometry imaging (MSI) data. Clustering methods are commonly employed to group similar pixels to form segmentation maps that describe different spatial features or regions of interest found in material samples. In this study, we initially assessed three clustering methods, <i>k</i>-means, DBSCAN, and SOM-RPM, for their ability to identify both intercluster and intracluster variability on a simple bivariate synthetic data set. The results showed that <i>k</i>-means struggles with nonconvex data, while DBSCAN requires arduous parameter tuning. SOM-RPM was found to be the most suitable for resolving complex data structures due to its topology preserving property. The application of SOM-RPM was further expanded to a specific data treatment task in MSI-background data removal. We used a complex microarray time-of-flight secondary ion mass spectrometry (ToF-SIMS) imaging data set as the exemplar, to investigate the impact of background removal on SOM-RPM models. Conspicuously, SOM-RPM produces robust segmentation maps, especially when the background was removed, that provide very useful insights into spectral diversity. We also incorporated a dimensionality reduction workflow by principal component analysis (PCA) and gained insights into the practicality of the method within ML pipelines. This study highlights the advantages of SOM-RPM in revealing underlying material properties and pipelining inefficiencies and measurement discrepancies to demonstrate its utility for applications such as forensic screening and analytical investigations.