Lectures and seminars Development of machine learning based tools for automatic volumetric mitochondria segmentation in 3D EM datasets

About the seminar

Recent advances in use of 3D electron microscopy techniques, such as focused ion beam scanning electron microscopy (FIB-SEM), allow for collection of large 3d EM datasets with nanometer resolution, enabling visualization of subcellular structures in great detail. A single cell may have over a thousand mitochondria and their morphology cannot be reasonably analyzed by hand without great effort, presenting a need for automated solutions.

Here we present a new FIB-SEM cochlear sensory hair cell mitochondria segmentation dataset which we used to train a novel, deep learning approach for automatic 3D mitochondria instance segmentation. Our fully automated mitochondria instance segmentation solution can accurately segment densely packed mitochondria which are often too large to fit in the neural network’s receptive field and is comparable in accuracy to our semi-automated, watershed segmentation approach.

Our tool is more time-efficient and calculates similar distributions for mitochondria volume and surface area in a fully unsupervised fashion. The use of this tool increases accessibility for analyzing mitochondria within large 3D electron microscopy stacks, enabling in-depth investigations of mitochondria morphology in normal and pathologic conditions.