EM-stellar: benchmarking deep learning for electron microscopy image segmentation.

Area of Science:

• Biophysics
• Computational Biology
• Microscopy

Background:

• Electron microscopy (EM) datasets have low contrast, hindering cellular ultrastructure segmentation.
• High-resolution EM techniques generate big datasets, increasing segmentation challenges.
• Deep learning (DL) offers automated segmentation but lacks benchmark analysis.

Purpose of the Study:

• To introduce EM-stellar, a Google Colab platform for benchmarking DL methods in EM data segmentation.
• To evaluate the performance of various state-of-the-art DL methods on diverse EM datasets.

Main Methods:

• EM-stellar is a Python-based platform available on GitHub under the MIT license.
• The platform allows users to benchmark DL segmentation performance on their own EM datasets.
• Performance evaluation across multiple metrics is integrated into the platform.

Main Results:

• DL method performance is dataset-dependent, influenced by image characteristics.
• No single DL method consistently outperforms others across all evaluation metrics.
• EM-stellar facilitates reproducible and comparative analysis of DL segmentation tools.

Conclusions:

• A standardized benchmarking approach is crucial for selecting appropriate DL methods for EM segmentation.
• EM-stellar provides a valuable resource for researchers in the field of EM data analysis.
• Future work may involve expanding the platform with more DL models and features.