Training deep learning models to classify nuclei in histopathology images usually depends on massive, expert-annotated datasets. Creating these datasets is costly and difficult to scale across different organs, which limits the development of robust models. A new arXiv preprint proposes a framework that aims to sidestep this bottleneck by using cell-level spatial transcriptomics as weak supervision.
The idea is to replace at least part of the manual labeling effort with information derived from spatial transcriptomics, which provides molecular profiles tied to cell positions. The authors argue that this can serve as a scalable source of supervision for nuclei classification, potentially reducing the need for exhaustive expert curation.
The abstract does not provide implementation details or experimental results, so the framework's effectiveness remains to be seen. Still, the direction is notable: it suggests an alternative path to training data that could make nuclei analysis more practical across diverse organ systems.