Abstract:

Extracting biological insight from microscopy images requires quantitative analysis methods. Many relevant analysis tasks operate on the level of individual cells or organelles, such as counting, morphological analysis, or tracking behavior over time. Manual cell-or-organelle-level analysis is prohibitively time-consuming at scale, creating the need for automated segmentation and related methods.Deep learning has consistently improved the quality of these methods over the past decade. Yet, until recently, it relied on bespoke training data for each task, creating a significant hurdle to practical adoption due to the large data annotation effort. The arrival of microscopy foundation models, such as μSAM developed by my group, has substantially improved the situation: these models work in many different conditions without the need for any further training data due to pre-training on very large datasets.

I will present our work on μSAM and our efforts to build μSAM2, which extends it to state-of-the-art multi-dimensional segmentation, cell tracking, and cell classification. I will also showcase how μSAM can substantially speed up image analysis across a variety of applications, from organelle segmentation in electron microscopy, to cell segmentation in large light-sheet microscopy volumes.