Stable Deep Neural Network Architectures for Mitochondria Segmentation on Electron Microscopy Volumes

Area of Science:

• Cell Biology
• Neuroscience
• Computational Biology

Background:

• Electron microscopy (EM) is crucial for identifying intracellular organelles like mitochondria.
• Deep learning models show promise for mitochondria segmentation but often lack reproducibility due to inaccessible code and training details.

Purpose of the Study:

• To conduct a comprehensive, reproducible comparison of state-of-the-art deep learning architectures for EM mitochondria segmentation.
• To establish best practices for model development and comparison in this field.

Main Methods:

• Implemented and evaluated various deep learning architectures, including U-Net variations, for mitochondria segmentation.
• Applied a standardized set of pre- and post-processing techniques across all models.
• Performed extensive hyperparameter sweeps and repeated runs to ensure stability and reliability.

Main Results:

• Identified stable architectures and training configurations that consistently achieve state-of-the-art performance.
• Outperformed previous methods on the EPFL Hippocampus, Lucchi++, and Kasthuri++ mitochondria segmentation datasets.
• Demonstrated the importance of reproducible research practices in deep learning for EM image analysis.

Conclusions:

• The study provides robust, reproducible deep learning models for mitochondria segmentation.
• Publicly available code and documentation facilitate further research and validation.
• Highlights the impact of standardized methodologies on achieving reliable and superior results in EM image analysis.