A Machine-Learning-Based Approach for Rapid 3D-Segmentation of cryo-Soft X-Ray Tomographic Datasets of Mammalian Cells
- Abstract number
- 165
- Presentation Form
- Poster Flash Talk + Poster
- DOI
- 10.22443/rms.mmc2021.165
- Corresponding Email
- [email protected]
- Session
- Stream 3: 3D+ Image Analysis
- Authors
- Michael Dyhr (3), Dr. Mohsen Sadeghi (1), Dr. Burcu Kepsutlu (3), Ralitsa Moynova (3), Dr. Stephan Werner (2), Prof. Dr. Gerd Schneider (2), Dr. James McNally (2), Prof. Frank Noé (1), Dr. Helge Ewers (3)
- Affiliations
-
1. Artificial Intelligence of the Sciences Group, Department of Mathematics and Computer Science, Free University of Berlin
2. Helmholtz-Zentrum Berlin für Materialien und Energie
3. Membrane Biochemistry Group, Department of Biology, Chemistry and Pharmacy, Free University of Berlin
- Keywords
cryo-soft x-ray microscopy, machine-learning, automated segmentation, artificial intelligence
- Abstract text
We developed a machine-learning-based approach that significantly reduces the time required for 3D-segmentation and data evaluation of cryo-soft X-ray tomograms of mammalian cells.
Cryo-soft X-ray tomography (cryo-SXT) is a powerful tool for elucidating the 3D ultrastructure of biological material down to a resolution of approximately 40 nm. Unlike fluorescence microscopy (FM) or electron microscopy (EM), the inherent contrast provided by soft X-ray imaging in the „water window“ permits the study of cellular architecture without additional labels or staining. Importantly, specimens can be quickly screened in the X-ray microscope for regions containing specific features of interest. Tomograms thereof can be acquired in a short time, covering a large volume and thus often capturing rare or previously uncharacterized cellular events or structures. These capabilities provide unique advantages to cryo-SXT compared to EM or FM.
Once a 3D cryo-SXT dataset has been acquired, it is often desirable to extract a 3D-model of cellular ultrastructure or to quantify cellular morphology. This requires accurate segmentation of the 3D volume which typically contains hundreds of slices. This data analysis is currently a key bottleneck in cryo-SXT. Segmentation of a single cryo-SXT dataset may require days or even weeks, since current automated segmentation tools are not very effective and instead manual segmentation is typically required. As a result, researchers have to invest a significant amount of time prior to any statistical analysis of their experimental data.
To overcome these limitations, we have developed a machine-learning approach to automatically segment cryo-SXT data. We trained a neural network with manual segmentations of representative cryo-SXT data obtained from vitrified mammalian cells. We find that the trained network automatically recognizes membrane structures in new, comparable 3D datasets, requiring as little as 10 minutes processing time using a single, modern GPU. This dramatic increase in the speed of the segmentation process will be important for overcoming current bottlenecks in cryo SXT by providing a solid basis for 3D-rendering and subsequent analysis of the 3D volumes.
In sum, our neural network approach has the potential to significantly accelerate the experimental workflow of researchers using cryo SXT, and in so doing make the method accessible to many more biologists.