Pseudo-spectral angle mapping to improve immune cell classification in highly multiplexed fluorescence microscopy images

Highly multiplexed fluorescence microscopy is an emerging technology that allows for spatial analysis of increasingly more classes of cells within human tissue—state-of-the-art methods are now probing up to 60 different protein markers within an image. This level of phenotypic resolution is ideal for uncovering the spatial underpinnings of immune cell interactions. However, defining cell types from this high-plex data is non-trivial. We present a method that borrows from hyperspectral image analysis to improve the accuracy and efficiency of immune cell classification in highly multiplexed fluorescence microscopy images. Treating the protein marker image channels as the spectral dimension of the images, we define reference "pseudospectra" representative of the ideal marker expression for all cell types of interest probed by the marker panel. Cosine similarity is computed for each reference pseudo-spectra to create class maps for each cell type in question. Features are extracted from these class maps—rather than the fluorescence images. We compare these methods to a decision-tree based classification method for classifying immune cells and unsupervised K-means clustering of mean pixel intensities across all image channels. We demonstrate that pSAM performs comparably, and potentially outperforms methods with similar levels of supervision.