Contrastive learning-driven framework for neuron morphology classification

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Neuron morphology classification is vital for understanding neural circuit function.
• Traditional methods face challenges with the complexity and diversity of neuronal shapes.
• Accurate classification is essential for advancing neuroscience research.

Purpose of the Study:

• To introduce PRT-net, a specialized network architecture for improved neuron morphology classification.
• To enhance classification performance and model generalization using novel techniques.
• To address data scarcity and imbalance issues in neuron morphology datasets.

Main Methods:

• PRT-net utilizes Complex Residual Structures and TreeLSTM for feature extraction.
• Innovative data augmentation strategies simulate morphological variations to boost robustness.
• A contrastive learning framework is employed to improve classification and generalization.

Main Results:

• PRT-net achieved high accuracies: 78.45% (BIL), 67.11% (JML), and 58.95% (ACT).
• Performance significantly surpassed current state-of-the-art methods on public datasets.
• Demonstrated substantial improvements of 2.9 and 3.3 percentage points on JML and ACT datasets.

Conclusions:

• PRT-net offers an efficient and robust solution for neuron morphology classification.
• The model shows strong adaptability to complex data distributions, validated by multiple metrics.
• This work advances automated analysis in neuroscience, facilitating deeper understanding of neural circuits.