SAM-EM Paper

SAM-EM (Segment Anything Model for Electron Microscopy) is a domain-adapted foundation model designed to enable real-time segmentation, tracking, and statistical analysis of liquid phase transmission electron microscopy (LPTEM) videos. Built on Meta’s Segment Anything Model 2 (SAM2), it is fine-tuned on curated synthetic LPTEM frames to overcome the low signal-to-noise ratio and temporal instability that limit existing segmentation methods. SAM-EM unifies segmentation and particle tracking with quantitative tools such as mean-squared displacement and trajectory distribution analysis, providing a full pipeline for extracting and interpreting nanoscale dynamics. Crucially, it maintains robustness under challenging conditions like 150 nm water thickness, the most experimentally relevant regime in LPTEM. By establishing a reliable, on-the-fly analysis framework, SAM-EM transforms LPTEM into a quantitative platform for single-particle tracking and accelerates data-driven materials discovery and design.