Related Experiment Video
Updated: May 15, 2025
12:08
From Voxels to Knowledge: A Practical Guide to the Segmentation of Complex Electron Microscopy 3D-Data
Published on: August 13, 2014
24.5K
Transfer learning improves performance in volumetric electron microscopy organelle segmentation across tissues.
Ronald Xie1,2,3,4, Ben Mulcahy5, Ali Darbandi5
1Terrence Donnelly Centre for Cellular & Biomolecular Research, University of Toronto, Toronto, ON, M5S 3E1, Canada.
Bioinformatics Advances
|April 8, 2025
Summary
Transfer learning significantly improves deep learning segmentation for volumetric electron microscopy (VEM) imaging. This approach reduces the need for extensive manual annotations, making nanoscale 3D biological structure identification more efficient.
Area of Science:
- Neuroscience
- Cell Biology
- Microscopy
Background:
- Volumetric electron microscopy (VEM) provides nanoscale 3D imaging of biological samples.
- Manual annotation of VEM data for organelle and cell identification is labor-intensive and time-consuming.
- Deep learning segmentation requires substantial labeled data, which is often unavailable for new VEM datasets.
Purpose of the Study:
- To develop and evaluate a transfer learning approach for deep learning-based segmentation in VEM.
- To reduce the manual annotation burden for segmenting biological structures in VEM datasets.
- To benchmark the proposed method against existing models and settings.
Main Methods:
- Pretraining deep learning models on diverse VEM datasets from multiple mammalian tissues and organelle types.
- Fine-tuning pretrained models on smaller, target-specific datasets.
- Benchmarking performance on published VEM datasets and a newly acquired rat liver dataset using serial block face scanning electron microscopy.
- Comparing the transfer learning approach with Segment Anything Model 2 and MitoNet in various settings (zero-shot, prompted, fine-tuned).
Main Results:
- The transfer learning method achieved high performance in segmenting multiple organelles.
- A relatively small amount of new training data was sufficient after pretraining.
- The rat liver dataset included a 56×56×11 m volume with detailed manual annotations for mitochondria and endoplasmic reticulum.
- Performance was benchmarked against other models, demonstrating the efficacy of the transfer learning strategy.
Conclusions:
- Transfer learning is an effective strategy to overcome data limitations in VEM segmentation.
- This approach significantly reduces the need for manual annotation, accelerating biological structure identification.
- The developed model and dataset are publicly available to facilitate further research in VEM analysis.
