Weakly Supervised Learning of 3D Deep Network for Neuron Reconstruction.

Area of Science:

• Neuroscience
• Computational Biology
• Image Analysis

Background:

• Accurate 3D neuron reconstruction is vital for understanding neural function and connectivity.
• Existing tracing methods struggle with noisy images and low-intensity neurites.
• Deep learning methods require extensive manual annotations, limiting their applicability.

Purpose of the Study:

• To develop a weakly supervised deep learning method for neuron reconstruction that eliminates the need for manual annotations.
• To improve the tracing of weak and low-intensity neurites in complex neuronal structures.
• To enhance the generalization capabilities of neuron reconstruction algorithms across diverse datasets.

Main Methods:

• Utilized a 3D residual convolutional neural network (CNN) for feature extraction.
• Implemented a weakly supervised learning framework with iterative pseudo-label refinement.
• Employed an automatic tracing method to generate initial pseudo-labels.
• Refined pseudo-labels by mining weak neurites based on tubularity and continuity.

Main Results:

• Achieved neuron reconstruction performance comparable to supervised CNN methods.
• Demonstrated effective detection of weak neurites in highly noisy images.
• Significantly improved tracing performance on both small and large-scale datasets (>100 GB).
• Outperformed several novel tracing methods on original, unannotated images.

Conclusions:

• The proposed weakly supervised method offers a robust solution for 3D neuron reconstruction from challenging microscopy images.
• Eliminating manual annotations enhances algorithm generalization and reduces labor intensity.
• The method shows high precision and effectiveness for large-scale neuroimaging datasets.