Related Experiment Video
Updated: Apr 11, 2026
10:25
Deep Learning-Based Segmentation of Cryo-Electron Tomograms
Published on: November 11, 2022
11.3K
NucleoNet and DropNet: Generalist deep learning models for instance segmentation of nuclei and lipid droplets from
Abhishek Bhardwaj1,2, Chris W Dell1,2, Melissa R Mikolaj1,2
1CCR Volume Electron Microscopy (CVEM), Center for Cancer Research, National Cancer Institute, National Institutes of Health, Bethesda, MD 20892, USA.
Biorxiv : the Preprint Server for Biology
|April 10, 2026
Summary
We developed NucleoNet and DropNet, deep learning models for segmenting nuclei and lipid droplets in electron microscopy images. These tools automate organelle segmentation, improving throughput for cellular imaging research.
Area of Science:
- Cell Biology
- Microscopy
- Bioinformatics
Background:
- Automating cellular organelle segmentation is crucial for high-throughput electron microscopy (EM) and volume EM (vEM) workflows.
- Deep learning (DL) has advanced segmentation, but model development is limited by a lack of training data for organelles beyond mitochondria.
- Manual annotation of complex EM images is time-consuming and labor-intensive.
Purpose of the Study:
- To crowdsource annotated datasets for nuclei and lipid droplets (LDs) from cellular EM images.
- To train and validate deep learning models for accurate instance segmentation of nuclei and LDs.
- To apply these models for quantitative analysis of cellular structures in different cancer models.
Main Methods:
- Crowdsourcing manual labeling of nuclei and LDs in complex cellular EM images.
- Training Panoptic DeepLab (PDL) models using large, heterogeneous annotated datasets and public vEM datasets.
- Developing instance segmentation models: NucleoNet for nuclei and DropNet for LDs.
- Evaluating model performance on diverse benchmarks.
Main Results:
- NucleoNet and DropNet achieved high-quality instance segmentation for nuclei and LDs across varied benchmarks.
- The models successfully quantified differences between 2D/3D in vitro cancer models and in vivo tumors.
- Demonstrated a pathway for robust quantitation in EM using automated segmentation.
Conclusions:
- NucleoNet and DropNet provide accurate and efficient automated segmentation of nuclei and LDs in EM images.
- These models facilitate quantitative comparisons of cellular structures in different biological contexts.
- Public availability via the empanada napari plugin enables widespread adoption for 2D and 3D EM image analysis.
