Few-shot learning for non-vitrified ice segmentation.

Area of Science:

• Structural biology
• Biophysics
• Cryo-electron microscopy

Background:

• Quantifying crystalline ice in cryo-electron tomography (cryo-ET) is crucial for accurate structural analysis.
• Existing methods lack adaptability and efficiency for diverse datasets.

Purpose of the Study:

• Introduce Ice Finder, a novel tool for crystalline ice quantification in cryo-ET.
• Apply meta-learning to unify diverse tomographic tasks within a single framework.
• Enhance domain generalization and rapid adaptation to new datasets using few-shot learning.

Main Methods:

• Developed Ice Finder, a tool leveraging the meta-learning paradigm.
• Utilized few-shot learning for improved domain generalization.
• Evaluated performance on in situ datasets from EMPIAR.

Main Results:

• Demonstrated the first application of meta-learning in cryo-ET ice quantification.
• Achieved rapid adaptation to new datasets with minimal examples.
• Showcased ease of use, fast processing, and millisecond inference times.

Conclusions:

• Ice Finder effectively quantifies crystalline ice in cryo-ET.
• The meta-learning framework provides a unified and adaptable approach.
• This tool significantly advances cryo-ET data processing and analysis.