Efficient data annotation for large-scale machine learning-based segmentations with webKnossos
- Abstract number
- 197
- Presentation Form
- Poster
- DOI
- 10.22443/rms.mmc2021.197
- Corresponding Email
- [email protected]
- Session
- Poster Session 1
- Authors
- Norman Rzepka (1)
- Affiliations
-
1. scalable minds GmbH
- Keywords
image segmentation, data annotation, 3d image data
- Abstract text
For many machine learning-based segmentation methods, high-quality data annotations remain crucial yet costly to acquire. We report a workflow for Machine Learning-based image analysis projects on large-scale datasets with efficient data annotation. For these workflows, we have developed webKnossos (Boergens et al. 2017), an open-source web-based tool for visualization, annotation, and sharing of large-scale 3D image datasets. webKnossos enables fast exploration of 3D datasets, sized up to petabytes, in the browser. Data can easily be shared between collaborators and annotators via links. webKnossos’ built-in annotation toolbox enables volume labeling, line-segment annotations (aka skeletons), and proof-reading.
In our workflow, training data is generated with the volume labeling tools. An integrated task management system enables distribution among several annotators. With the obtained ground truth data, we can train a neural network model. We can visualize the predictions overlaid onto the raw data using the multi-layer feature. After deriving a segmentation from the model predictions with a watershed algorithm, segments can be visualized as 3D meshes for visual inspection. Errors can be corrected with the proof-reading tools. Next, we evaluate our segmentation. The skeleton annotation tool enables gathering evaluation data (e.g., neuron paths or locations of nuclei). For the case of tracing evaluation neurons in volume electron microscopy (EM) data, we utilize the flight mode of webKnossos, a fast egocentric annotation mode. At all stages, data may be imported and exported for interoperability with other tools.
In my talk, I will focus on the use case of neuron reconstruction from volume EM data. However, webKnossos is well-suited for 3D datasets of multiple modalities including EM, fluorescence microscopy, X-ray CT, and MRI.
https://webknossos.org
- References
Boergens, K., Berning, M. et al. "webKnossos: efficient online 3D data annotation for connectomics". Nat Methods 14, 691–694 (2017). https://doi.org/10.1038/nmeth.4331