Current Progress and Challenges in Large-Scale 3D Mitochondria Instance Segmentation.

Area of Science:

• Neuroscience
• Biomedical Imaging
• Computational Biology

Background:

• Accurate 3D instance segmentation of mitochondria is crucial for understanding cellular function.
• Previous datasets for mitochondria segmentation were limited in scale, hindering progress.

Purpose of the Study:

• To benchmark 3D instance segmentation methods for mitochondria in large-scale electron microscopy datasets.
• To evaluate and improve the scoring system for mitochondria segmentation challenges.

Main Methods:

• Organized the MitoEM challenge with large-scale 3D electron microscopy datasets from human and rat cortex.
• Collected and presented eight top-performing participant methods and baseline strategies.
• Corrected ground truth annotations and retrospectively evaluated the challenge scoring metric.

Main Results:

• Eight top-performing segmentation methods were presented, with several outperforming baseline strategies.
• Identified limitations in the original scoring metric, specifically regarding false positives and size-based grouping.
• Proposed a new scoring system to more accurately assess mitochondria segmentation correctness.

Conclusions:

• Substantial progress was made in mitochondria segmentation, but challenges remain for complex morphologies.
• The proposed scoring system offers a more robust evaluation of segmentation performance.
• The MitoEM challenge dataset and platform remain open for continued research and development.