Unsupervised Semantic Segmentation Models for Region of Interest Identification

Spatial omics technologies, such as mass spectrometry imaging (MSI), can capture biomolecular distributions and their spatial locations directly from a tissue, but these distributions are not easily associated with tissue morphology without additional microscopy performed on the sample. To identify tissue regions of interest (ROIs), a reference image (such as a stained tissue, microscopy images, etc.) may be used and paired with mass spectrometry data. Due to the intense time requirements of manually labeling ROIs, segmentation models have been demonstrated as time-saving alternatives, as they automatically annotate ROIs by grouping like pixel intensities together. Supervised approaches trained on specific image types are preferred, but when those cases are not available, generalizable unsupervised segmentation models may then be used. Here, eight unsupervised semantic segmentation algorithms in R and Python, representing both statistical and machine learning algorithms, were compared for their ability to match manual annotations of 30 tiled PAS-stained kidney images and 25 tiled plant root images. Noise reduction techniques such as dimension reduction were tested to see whether they improved segmentations, and all models were applied to the full stitched images. Performance metrics were calculated to provide recommendations on the highest performing models to demonstrate their potential for automated annotation of tissue ROIs. At small cluster sizes, k-means and pytorch-tip tended to have the best performance in terms of balanced accuracy and time, though all algorithms had decreased performance at higher cluster numbers. Lastly, the segmentation model choice was demonstrated to have an impact on downstream statistics, highlighting the importance of testing and selecting the best segmentation model on a case-by-case basis, as no one model had the best performance in every comparison.