Instance segmentation of mitochondria in electron microscopy images with a generalist deep learning model trained on a diverse dataset
Problem
Automatically annotating each mitochondrion accurately in any 2D or volume electron microscopy image was an unsolved computational challenge.
Why it matters
Accurate organelle annotation at scale is needed for cell biology research but is infeasible to do manually.
What the AI did
MitoNet, a deep learning model, was trained on a heterogeneous dataset of ~1.5×10^6 unlabeled cellular EM images with ~135,000 segmented mitochondrial instances, and generalizes to unseen volume EM datasets.
Results
Segmented ~135,000 mitochondrial instances; performs well on challenging benchmarks and previously unseen volume EM datasets containing tens of thousands of mitochondria
Impact
Released empanada Python package and napari plugin enabling rapid inference and proofreading; transparent peer review record included.
Record
| Venue | Cell Systems |
|---|---|
| Published | 2023 |
| Work type | article |
| Authors | Ryan Conrad, Kedar Narayan |
| Citations | 113 (OpenAlex, at import time) |
| References | 80 |
| Model or method | MitoNet |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1016/j.cels.2022.12.006 |
| Open access | hybrid |
| Keywords | mitochondria, instance segmentation, electron microscopy, deep learning, MitoNet, napari |
Importance rating
significant — thousands to millions affected
Automatic organelle annotation is an unsolved challenge enabling cell biology research at scale; abstract shows generalization across large EM datasets but frames impact as a research capability.