Model Background#
The empanada-napari plugin is built to democratize deep learning image segmentation for researchers in electron microscopy (EM). It ships with MitoNet, a generalist model for the instance segmentation of mitochondria.
MitoNet#
MitoNet is an generalist deep learning model specifically developed for instance segmentation of mitochondria within EM images. It is pre-trained on a diverse dataset of approximately 1.5 million unlabeled cellular EM images, and subsequently trained on ~135k labeled mitochondrial instances.
Leveraging elements from Panoptic DeepLab and Panoptic BiFPN models, MitoNet’s architecture is optimized to handle the complex shapes of mitochondria in EM images. The model’s performance was evaluated on challenging volume EM benchmarks to test its ability for accurate and precise segmentations. (Check out the full paper linked below)
These models accept a grayscale EM image and output a semantic segmentation, up-down and right-left offsets, and a heatmap with peaks at object centers. After postprocessing, a panoptic (or in the case below, instance) segmentation is created.
Check out the full paper in Cell Press here: