A Generalizable AI-Based Segmentation for Morphological Single Cell- and Sub-cellular Profiling

Biomolecular Imaging and Informatics 2024 - SBI2 11th Annual Conference
September 18, 2024

In high content screens, quantification of outcomes typically relies on identifying and quantifying distinct features of the cells images, which in turn requires segmentation of cells in e.g. cytosol vs nucleus, or even more fine-grained segmentation in subcellular structures.
Here, we detail two workflows using the foundational Segment Anything Model (SAM) combined with prompt engineering that introduces biologically informed constraints. This approach introduces a generalizable segmentation workflow applicable to a wide variety of cellular phenotypes, enabling automated extraction of morphological features at both cellular and subcellular levels. We illustrate its robustness on diverse datasets, including a receptor internalization assay and images from the Human Protein Atlas.
Our study shows that prompt engineering can eliminate the need for retraining and demonstrates the potential of segmentation foundation models in terms of enablement for rapid prototyping and for future routine applications.