Efficient spatial segmentation of large imaging mass spectrometry datasets with spatially aware clustering

Motivation: Imaging mass spectrometry (IMS) is one of the few measurement technology s of biochemistry which, given a thin sample, is able to reveal its spatial chemical composition in the full molecular range. IMS produces a hyperspectral image, where for each pixel a high-dimensional mass spectrum is measured. Currently, the technology is mature enough and one of the major problems preventing its spreading is the under-development of computational methods for mining huge IMS datasets. This article proposes a novel approach for spatial segmentation of an IMS dataset, which is constructed considering the important issue of pixel-to-pixel variability. Methods: We segment pixels by clustering their mass spectra. Importantly, we incorporate spatial relations between pixels into clustering, so that pixels are clustered together with their neighbors. We propose two methods. One is non-adaptive, where pixel neighborhoods are selected in the same manner for all pixels. The second one respects the structure observable in the data. For a pixel, its neighborhood is defined taking into account similarity of its spectrum to the spectra of adjacent pixels. Both methods have the linear complexity and require linear memory space (in the number of spectra). Results: The proposed segmentation methods are evaluated on two IMS datasets: a rat brain section and a section of a neuroendocrine tumor. They discover anatomical structure, discriminate the tumor region and highlight functionally similar regions. Moreover, our methods provide segmentation maps of similar or better quality if compared to the other state-of-the-art methods, but outperform them in runtime and/or required memory.