Semi-Supervised Neuron Segmentation via Reinforced Consistency Learning

Area of Science:

• Neuroscience
• Computer Vision
• Machine Learning

Background:

• Deep learning methods have advanced neuron segmentation in electron microscopy (EM).
• Current methods require extensive manual annotations, which are costly and time-consuming.
• Limited annotations lead to fragile and unreliable segmentation models.

Purpose of the Study:

• To develop a semi-supervised learning method for neuron segmentation that effectively utilizes unlabeled EM data.
• To overcome the limitations of data-hungry supervised learning approaches in EM neuron segmentation.
• To improve the generalizability and robustness of neuron segmentation models with scarce labeled data.

Main Methods:

• A two-stage semi-supervised learning framework is proposed for neuron segmentation.
• Stage 1: Network pre-training using a proxy task of reconstructing original volumes from perturbed versions.
• Stage 2: Regularizing supervised learning with pixel-level prediction consistency on unlabeled data and their perturbed counterparts.

Main Results:

• The proposed method demonstrates superior performance compared to purely supervised learning, especially with limited labels.
• Achieved up to a 400% gain in the Volume of Interest (VOI) metric with minimal labeled data.
• Performance is comparable to supervised models trained on ten times the amount of labeled data.

Conclusions:

• The reinforced consistency learning strategy effectively extracts information from unlabeled EM data.
• The method significantly enhances neuron segmentation accuracy and generalizability in low-label scenarios.
• This approach offers a more efficient and scalable solution for large-scale EM data analysis.