Medical image analysis is constrained less by algorithms than by the cost of expert annotation. I will describe an alternative paradigm in which segmentation emerges from the geometry of pretrained neural representations rather than from labelled training. The pipeline embed, project with UMAP, cluster via the graph Laplacian is simple, modular, and requires only light post-hoc labelling. I will focus on HistoPAINT, our histopathology application, and close with thoughts on where geometric, low-supervision pipelines are likely to displace fully supervised ones in scientific imaging.