🟢 Mitochondrion

easymode segment mitochondrion

The mitochondrion model was trained at 30 Å/px and has a receptive field of almost 5000 Å. As a result, it is mostly able to recognize the full extent of mitochondria, and at high magnifications can be applied to tomograms in one go - no sliding window required, and thus utilizing the global context of the tomogram. The model was trained on manually curated 2D Ais UNet generated pseudolabels, using training data from human, mouse, baker's yeast, fission yeast, chlamydomonas, dictyostelium discoideum, and a number of eukaryotic species, as well as on various prokaryotic species to include counterexamples to mitochondria. We expect the model to perform well on any other FIB-milled cultured cell type, but note that mitochondria can have very different morphologies in different tissue samples; for example in human sperm (also included in the training data) - and advise caution in such cases.

Validation by subtomogram averaging of mitochondrial complexes - using the mitochondrion model to masks picks of ribosomes and various mitochondrial complexes - is planned.

Example output

Example of easymode segment mitochondrion output overlaid on a tomogram from EMPIAR-11899 (FIB-milled D. discoideum), a dataset which was not used to train this model.