From Masks to Morphology: Features & Self-Supervised Embeddings

The problem — A segmentation mask is not a measurement. The features stage turns each object into the vector a biologist actually reasons over, and that representation decides what phenotypes you can ever detect. Pick it carelessly and the assay is blind to exactly the differences you ran it to find.

What it is / how it works — Two families. Classical profiling measures hundreds of named morphology, texture, and intensity descriptors per object — the CellProfiler lineage that powers most Percent Replicating checks today. The features are auditable: every column has a name and a biological reading. The learned family instead pretrains an encoder to emit an embedding. cytoself does this with no labels at all — a VQ-VAE trained on a protein-identity pretext task that out-clusters CellProfiler on organelle and complex membership, the self-supervised analog of morphological profiling. DeepProfiler and scDINO carry the same idea onto the self-supervised ViT line that DINOv2 generalized: frozen, transferable features learned without annotation.

Where it breaks — Embeddings buy sensitivity and lose a name. A learned axis that separates two compounds tells you that they differ, not how — and it will just as happily separate two scanners or two plate batches. Self- supervised models are notorious for encoding acquisition nuisance; cytoself's own authors flag residual batch effects they do not suppress. So the features stage never stands alone: whatever vector you ship still has to clear correction and a reproducibility gate before anyone trusts it, which is the whole argument behind Why Microscopy Needs Verified Outputs.