A Semi-supervised Pipeline for Accurate Neuron Segmentation with Fewer Ground Truth Labels

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Two-photon calcium imaging allows recording activity from thousands of neurons at cellular resolution.
• Analyzing large-scale neural recordings necessitates automated neuron segmentation methods.
• Current supervised methods require extensive manually generated ground truth labels.

Purpose of the Study:

• To develop a semi-supervised pipeline to reduce the number of training labels for neuron segmentation.
• To improve the efficiency of analyzing large neural recordings from calcium imaging data.
• To achieve state-of-the-art accuracy in neuron segmentation with reduced manual annotation effort.

Main Methods:

• A semi-supervised pipeline utilizing neural network ensembling to generate pseudolabels.
• Training a shallow U-Net model using the generated pseudolabels.
• Validation on three publicly available two-photon calcium imaging datasets.

Main Results:

• The proposed method outperformed existing segmentation methods when trained on limited ground truth labels.
• Achieved state-of-the-art accuracy using approximately 25% of the labels required by fully supervised methods.
• Demonstrated superior accuracy compared to state-of-the-art methods when trained on extensive labels.

Conclusions:

• The semi-supervised pipeline effectively reduces the manual effort required for neuron segmentation.
• Enables accurate processing of large neural recordings, accelerating neuroscience research.
• Minimizes the time and resources needed for generating training data for automated segmentation.