QuantEM

Electron microscopy resolves cellular ultrastructure at nanometer scale, but turning those images into measurements depends on segmentation, and segmentation is still largely manual — annotating one acquisition can consume days to months. Tools built to automate it are usually trained on a single laboratory's data and degrade elsewhere, because public EM data is sparse and existing annotations cover a narrow slice of tissues, species, and protocols.

QuantEM, from the Arrojo e Drigo laboratory at Vanderbilt University, addresses the data problem first and the model second. The authors crawled eight public repositories — among them OpenOrganelle, EMPIAR, BossDB, and the BioImage Archive — screened entries with an LLM-based metadata filter, then emailed 808 corresponding authors whose published EM images had never been deposited anywhere; 27 groups contributed. With unpublished in-house acquisitions, this released 3,249 new assets across 154 datasets and widened the public EM corpus by roughly 30%. The assembled corpus spans 16,107 two-dimensional images and 1,720 three-dimensional acquisitions from 101 species and 339 distinct tissues.

On that corpus the team pretrained an EM-native vision transformer and built a platform around it: segmentation heads for mitochondria, endoplasmic reticulum, nuclei, and lipid droplets, with interactive proofreading, guided fine-tuning, and morphometrics in a standalone application and a napari plugin. A second family puts the same heads on the OmniEM EM-DINO encoder, trading parameter count against accuracy per organelle. Where MitoStructSeg goes deep on one organelle's internal architecture, QuantEM goes wide across organelles, tissues, and protocols.

#Key Features

• Four organelles, two encoder sizes: Eight checkpoints cover mitochondria, ER, nuclei, and lipid droplets on either an 86M-parameter ViT-B/16 or the 302M-parameter OmniEM ViT-L/14, so a lab can choose per organelle and per GPU budget.
• Label-efficient adaptation: Fine-tuning updates lightweight decoder heads rather than the backbone, handling custom imaging conditions from a handful of manual annotations at a fraction of full-model training cost.
• A published, attributed data corpus: Every source dataset is enumerated with a DOI or repository URL in companion tables, and outreach images were deposited into the BioImage Archive rather than only into the training set.
• Proofreading and morphometrics: The napari plugin wraps segmentation, correction, batch inference, head-only fine-tuning, and quantitative analysis in one interface for users who do not work in Python.

#Technical Details

The QuantEM encoder is a ViT-B/16 with 86 million parameters trained from scratch under the DINOv3 self-supervised recipe. Assets are tiled into minimally overlapping 2048x2048 crops, capped at 400 tiles per asset so large volumes cannot dominate, then randomly downscaled and cropped to a 512-pixel context window; quality plateaued near 650,000 training steps. Candidate encoders were compared with frozen backbones and identical UPerNet decoders on ER and mitochondria. Across ablations of necks, decoders, losses, input scaling, and test-time adaptation, backbone adaptation dominated every other design choice: LoRA suited the ViT-L models, while the ViT-B gained more from unfreezing its last four transformer blocks or full fine-tuning.

Across held-out sources the best configurations reach mean Dice scores of 0.725 for nuclei, 0.630 for mitochondria, 0.577 for lipid droplets, and 0.453 for ER, reflecting how difficult thin curvilinear membrane networks remain. Reconstructing published train/test splits in the same framework, QuantEM is competitive with OrgSegNet on its native plant dataset and substantially outperforms DeepContact on its native COS-7 data. On immuno-electron microscopy, where preparation sacrifices membrane contrast for antigen accessibility, fine-tuning on two annotated crops raises QuantEM to approximately 87% Dice; MitoNet required far more labels, never matched that accuracy, and overfit as labels were added.

#Applications

QuantEM targets cell biologists and imaging core facilities generating 2D electron micrographs who need organelle composition, morphology, and spatial relationships quantified faster than manual annotation allows. Its demonstration case is instructive: combining automated organelle masks with correlative immuno-EM in mouse liver, the authors assigned gold-labeled glucokinase molecules to compartments and showed that cytoplasmic glucokinase is enriched inside mitochondria and within 100 nm of a mitochondrial border relative to Monte Carlo randomized distributions — an analysis that would otherwise have required extensive hand segmentation.

#Impact

The most durable contribution may be the corpus rather than the checkpoints: cataloging fragmented holdings and persuading laboratories to deposit unpublished images enlarged the shared substrate every future EM segmentation model trains on. The benchmarking is also unusually explicit about which architectural decisions matter. Constraints remain: the models are two-dimensional only, cover four organelles, and stay well below expert manual segmentation, most visibly on ER; the authors attribute this to biological heterogeneity and scarce annotation rather than limited receptive field. The preprint has not been peer reviewed. Weights and source tables are published on Hugging Face under CC BY 4.0, while the quantem-core library and napari-quantem plugin that load them are documented on the model card but not yet posted to a public package index or code repository.