Diffusion models for nuclei segmentation in low data regimes

Overview & Synopsis

Nuclei detection and characterization in histopathological tissue assessment is of utmost importance for different clinical workflows, such as the characterization of tumor micro-environments. Utilizing robust computational models for such a task could allow streamlining the process. However, obtaining accurate segmentation maps for histopathological slides can be quite tedious and expensive, while also being subject to inter/intra-reader variability. Learning robust and precise segmentation models from only a small amount of data would be therefore a very interesting alternative to fully supervised methods relying on a huge amount of annotations. Inspired by diffusion models' recent advances, this paper proposes a method for obtaining nuclei segmentation maps under low data regimes. In particular, diffusion models are used for learning powerful pixel-level representations of digital pathology patches that could require only a few amounts of annotated data to provide multiclass segmentation maps of different nuclei. Various insights about the use of these models for the representation of digital pathology patches are provided. Comparisons with other self-supervised and fully supervised methods highlight the advantages of the use of these models for nuclei segmentation.