Recent technical advances in volume electron microscopy (vEM) and artificial intelligence-assisted image processing have facilitated high throughput quantifications of cellular structures, such as mitochondria that are ubiquitous and morphologically diversified. A still often overlooked computational challenge is to assign cell identity to numerous mitochondrial instances, for which both mitochondrial and cell membrane contouring used to be required. Here, we present a vEM reconstruction procedure (called mito-segEM) that utilizes virtual path-based annotation to assign automatically segmented mitochondrial instances at the cellular scale, therefore bypassing the requirement of membrane contouring. The embedded toolset in webKnossos (an open-source online annotation platform) is optimized for fast annotation, visualization, and proofreading of cellular organelle networks. We demonstrate broad applications of mito-segEM on volumetric datasets from various tissues, including the brain, intestine, and testis, to achieve an accurate and efficient reconstruction of mitochondria in a use-dependent fashion.
The deep learning framework powered by PyTorch for automatic and semi-automatic semantic and instance segmentation in connectomics was provided by the Visual Computing Group (VCG) at Harvard University.
Refer to the Pytorch Connectomics wiki, specifically the installation page, for the most up-to-date instructions on installation on a local machine or high-performance cluster.
WEBKNOSSOS is open-source so you can install it on your server. Check out the documentation for a tutorial on how to install WEBKNOSSOS on your server.
pytorch, webknossos, fastremap, h5py, tifffile, zarr
The pre-trained mitochondrial segmentation model (Lu et al., 2024) was based on a residual 3D U-Net architecture with four-down/four-up layers, which was provided by PyTorch Connectomics. The model was trained to classify each voxel of the input stack (17 consecutive 256 × 256 pixel-sized images) into the “background”, “mitochondrial mask”, and “mitochondrial contour” categories. The model output was a two-channel image stack with the same format as the input, including the predicted probability maps of mitochondrial masks and contours.
To generate mitochondrial instance masks, the seeds of mitochondria (or markers) were determined with a high mask probability and low contour probability by thresholding. Then, the marker-controlled watershed transform algorithm (part of the scikit-image library) was employed to generate high-quality instance masks of mitochondria with the seed locations and the predicted probability map of the masks.
The segmentation of mitochondria was imported into the webKnossos using Python scripts.
python segmentation_to_webknossos.py
In the “toggle merger mode” and with the option “hide the unmapped segmentation” selected, a start point was seeded and associated mitochondria were annotated one after another through the mouse right-clicks within individual instances. Upon each valid assignment, the corresponding mitochondrial instance would become visible with a pseudo-color and linked by an active node, so that missing and multiple annotations of mitochondrial instances could be minimized. Note that the “toggle merger mode” does not allow a mouse click outside the segments and ignores redundant annotations of a single segment. Finally, the assembly of the nodes was utilized to specify the associated mitochondrial instances that could be then operated as a defined group with i.e. self-written Python scripts.
First, you open the dataset, then click "Create Annotation". Switch to skeleton mode and create a new tree. Then, you can follow the steps above.
Mouse brainstem dataset, Mouse intestine dataset, Mouse testis dataset
You can download the .nml file of the mitochondrial virtual path by the link.
The nodes of mitochondria are based on virtual path annotation, which could be extracted from webKnossos.
python extract_virtual_path_node_info.py
All mitochondrial volume was precomputed.
python volume_precomputation.py
The mitochondrial complexity index(MCI) was computed.
python compute_MCI.py
Here, we used Amria software to 3D render the reconstruction of mitochondria in various cells.
.
This project is built upon numerous previous projects. Especially, we'd like to thank the contributors of the following GitHub repositories:
- pytorch_connectomics. Visual Computing Group (VCG) at Harvard University
- webKnossos. Scalable minds GmbH, Potsdam, Germany
For a detailed description, please read this paper. If you use the method in your research, please cite:
Jiang et al., 2024. Efficient cell-wide mapping of mitochondria in electron microscopic volumes using webKnossos. BioRxiv.