SELF-SUPERVISED LEARNING TO IMPROVE TOPOLOGY-OPTIMIZED AXON SEGMENTATION AND CENTERLINE DETECTION.

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Preserving axonal topology is crucial for understanding brain connectivity and regional intersections in large-scale brain mapping.
• Developing algorithms using unannotated data can significantly reduce the need for time-consuming expert annotations.

Purpose of the Study:

• To develop and evaluate a self-supervised learning framework for accurate axon segmentation and centerline detection in mouse brain data.
• To leverage pretrained models for improved feature representation and topological preservation in neural structures.

Main Methods:

• Applied self-supervised training to a Residual 3D U-Net with an auxiliary classifier for voxel sample ordering.
• Utilized pretrained encoder weights to train a Residual 3D U-Net with a topology-preserving loss function (soft centerline-Dice) for segmentation and centerline detection.
• Investigated the sensitivity of topological loss function hyperparameters (α and k) for axon analysis.

Main Results:

• Achieved improved feature space representation of axonal structures compared to previous methods.
• Demonstrated enhanced evaluation metrics for axon segmentation and centerline detection.
• Identified sensitivity of topological loss hyperparameters to the specific tasks of axon segmentation and centerline detection.

Conclusions:

• Self-supervised learning effectively enhances axon segmentation and centerline detection in large-scale brain mapping.
• The proposed method improves topological preservation and feature representation of neural structures.
• Hyperparameter tuning of the topologically-aware loss function is critical for optimizing performance in axon analysis.