MitoNet automatic instance segmentation of mitochondria in the OpenOrganelle Mouse Kidney dataset

MitoNet is a generalist model for the instance segmentation of mitochondria in electron microscopy images. This dataset is the automatically generated result from applying MitoNet to the large OpenOrganelle Mouse Kidney FIB-SEM dataset at 16 nm voxel size. It contains nearly 220,000 segmented objects stored in a chunked Zarr array with chunk size of (512, 512, 512).

This version of MitoNet predicted both the semantic segmentation, to distinguish mitochondria from background, and contours of each instance. Predictions were hardened at a confidence threshold of 0.5 for the semantic segmentation and 0.3 for the contour segmentation. Connected components and the watershed algorithm were then applied to create an instance segmentation. Instances that were found to be smaller than 4,000 voxels were removed as likely false positives.

Inference took 4 hours on a single Biowulf compute node equipped with a A100 GPU, 64 GB RAM, and 32 CPUs cores.

The data file contains a Zarr directory that can be loaded in Python for further analysis or Napari for visualization.