Nuc2Vec: Learning Representations of Nuclei in Histopathology Images with Contrastive Loss

The tumor microenvironment is an area of intense interest in cancer research and may be a clinically actionable aspect of cancer care. One way to study the tumor microenvironment is to characterize the spatial interactions between various types of nuclei in cancer tissue from H&E whole slide images, which requires nucleus segmentation and classification. Current methods of nucleus classification rely on extensive labeling from pathologists and are limited by the number of categories a nucleus can be classified into. In this work, leveraging existing nucleus segmentation and contrastive representation learning methods, we developed a model that learns vector embeddings of nuclei based on their morphology in histopathology images. We show that the embeddings learned by this model capture distinctive morphological features of nuclei and can be used to group them into meaningful subtypes. These embeddings can provide a much richer characterization of the statistics of the spatial distribution of nuclei in cancer tissue and open new possibilities in the quantitative study of the tumor microenvironment.