Abstract
Mitochondria are extremely pleomorphic in biology. Automatically identifying each one precisely and accurately from any 2D or volume electron microscopy (volume EM) dataset is an unsolved computational challenge. Current deep learning (DL) models are trained within limited contexts, restricting their widescale utility and potential as a universal or generalist solution for mitochondrial segmentation. To address this, we amass a highly heterogeneous ~1.5 x 106 unlabeled cellular EM image dataset and a ~22,000, partially crowdsource-labeled, mitochondrial instance segmentation dataset. We release MitoNet, a DL model trained on these data, which performs well on new and challenging volume EM benchmarks. An accompanying Python package and napari plugin, called empanada, can be used for efficient training, inference, and clean-up of instance segmentations on EM images.
Competing Interest Statement
The authors have declared no competing interest.