EMCF ecosystem: Towards pretrained foundation model for electron microscopy image analysis

Abstract

Volume electron microscopy (vEM) enables nanoscale visualization of three-dimensional (3D) cellular ultrastructure, providing critical insights into physiological processes and pathological alterations. However, its application to large-scale biological tissues remains constrained by two major bottlenecks: prolonged image acquisition and inefficient data processing. Here, we present EMCF ecosystem (EMCFsys), an integrated ecosystem designed to overcome these challenges through three key components: a large-scale benchmark dataset (EMCFD) comprising 4,002,802 high-quality images across 14 EM modalities and 6 biological kingdoms; a foundation image restoration model (EMCellFiner); and a scalable image analysis foundation model (EMCellFound). Together, these modules systematically enhance image quality and substantially improve analysis efficiency. Our results show that EMCellFiner outperforms specialist models in restoring degraded images, even surpassing original ground truth sharpness in certain artifact regions, and reduces imaging time by 16-fold by enabling low-resolution and low dwell time acquisition. EMCellFound exhibits exceptional feature discriminability, outperforms specialist models in classification, semantic segmentation and instance segmentation. It also enables high-precision 3D reconstruction of organelles (e.g., endoplasmic reticulum) with minimal labeled data (0.01% of total volume). We validated the EMCFsys on unseen datasets across diverse biological contexts and imaging platforms. By publicly releasing both the dataset and models, we establish a scalable paradigm for automated, high-throughput vEM data interpretation, accelerating exploration of lifes nanoscale structure and function across biology.