A distribution-aware semi-supervised pipeline for cost-effective neuron segmentation.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Semi-supervised learning (SSL) is cost-effective for neuron segmentation in electron microscopy (EM) volumes.
• SSL uses unlabeled data to improve supervised training for neuron boundary prediction.
• Distribution mismatch between labeled and unlabeled data limits SSL model generalization in EM neuron segmentation.

Purpose of the Study:

• To develop a distribution-aware pipeline to enhance semi-supervised neuron segmentation in EM volumes.
• To address the distribution mismatch issue inherent in SSL for EM data.
• To improve the generalization capabilities of neuron segmentation models.

Main Methods:

• Data-level selection of representative sub-volumes for annotation using unsupervised distributional similarity.
• Model-level encouragement of consistent predictions across mixed views of labeled and unlabeled data.
• Network design to align feature distributions and learn shared semantics.

Main Results:

• The developed pipeline effectively enhances semi-supervised neuron segmentation in EM volumes.
• The method demonstrates improved model generalization on diverse EM datasets.
• The approach successfully addresses the distribution mismatch problem.

Conclusions:

• The distribution-aware pipeline significantly improves semi-supervised neuron segmentation in EM.
• This method has the potential to reduce manual proofreading efforts.
• The approach can accelerate large-scale connectomic reconstruction.