Crowdsourcing the creation of image segmentation algorithms for connectomics

Area of Science:

• Neuroscience
• Computational Biology
• Image Analysis

Background:

• Automating neural circuit reconstruction from electron microscopic (EM) images is crucial for understanding brain function.
• Manual reconstruction is time-consuming and labor-intensive.
• Standardized challenges are needed to drive progress in automated methods.

Purpose of the Study:

• To organize the first international challenge on 2D segmentation of brain EM images.
• To evaluate and advance automated methods for neural circuit reconstruction.
• To identify limitations in current segmentation and scoring approaches.

Main Methods:

• Organized a 2D segmentation challenge for brain EM images.
• Participants submitted predicted boundary maps.
• Scoring based on agreement with human expert annotations.
• Utilized convolutional neural networks (deep learning) for segmentation.

Main Results:

• The winning team, new to EM images, used a convolutional network (deep learning).
• Deep learning has become a standard for EM image segmentation.
• The challenge appears saturated due to inherent 2D scoring ambiguities and dataset size.
• The scoring system lacked robustness to neurite border width variations.

Conclusions:

• Deep learning is highly effective for 2D EM image segmentation in neuroscience.
• Current 2D segmentation challenges may have reached their limits.
• A more robust scoring system is needed, particularly for future 3D segmentation challenges.