Electron Tomography Segmentation
Cryo-electron tomography (cryo-ET) is a powerful imaging technique used to study the three-dimensional (3D) cellular components and macromolecular complexes in their near-native state, providing insights into their organization and interactions. The data acquired can be noisy, and the sample may undergo distortions during the imaging process.
Using Machine Learning techniques U-Net and Autoencoder, we are able to successly segment the cellular structures like ribosome, membrane, microtubules, filament.
Our Contributions:
- Developed a Multi-UNet architecture to improve the performance of multi-class segmentation by 15%
- Added different strategies to effectively train the deep learning U-Net architectures from a sparsely labelled tomogram images.
- Achieved 92% accuracy in segmenting electron structures like ribosomes, membrane using <1% of the entire tomogram for training
- Designed U-NeXt architectures, combining the ConvNeXt and U-Net, specifically tailored for tomograms captured at different scales, resulting in a f1 score of 85% for segmentation