The AI features in WEBKNOSSOS are now production-ready, and we tested them on a real-world volume EM dataset of mouse cortex (Motta et al., 2019). We trained a custom instance segmentation model and deployed it across the full 140 GB volume (5445 × 8380 × 3285 voxels; 61.2 µm × 94.2 µm × 92.0 µm) in just a few clicks.
Ground truth generation
Starting with 20 bounding boxes, we annotated mitochondria in WEBKNOSSOS using the Quick Select tool in 3D. The prediction depth was set to match the height of the bounding boxes so the tool would automatically segment mitochondria throughout, i.e. the Quick Select tool predicted 32 sections along the Z-axis. Whenever the Quick-select tool didn’t perform well enough, we used the manual brush and eraser tools, combined with the interpolation tool.

Annotating mitochondria in WEBKNOSSOS with the Quick-select tool
Ground truth annotation took about 4 hours of work in total, while training and testing the model was fast and effortless.
Model testing and iterating
After testing the model on a small bounding box, we observed three types of errors:
● myelin was misclassified as mitochondria
● dense vesicle-filled synaptic regions were sometimes also mistaken for mitochondria
● some mitochondria were merged

Errors in the test segmentation of the first model version
To fix these, we annotated a few additional boxes covering these cases and retrained the model. After two iterations, the segmentation results on the test bounding boxes were highly accurate.
Full volume segmentation
Finally, we applied the model to the full dataset, producing high-quality mitochondria segmentation across the entire volume. The inference lasted a few hours.

Scrolling through the results, inspecting a few 3D meshes
Check out the results here and explore beautiful mitochondria.
Get started with your own dataset using this tutorial and guides on sampling bounding boxes and creating high-quality ground truth annotation.