Multi-level feature fusion network for neuronal morphology classification.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Bioinformatics

Background:

• Neuronal morphology is crucial for understanding neural function.
• Existing methods often use single or concatenated features, limiting classification performance.
• Complementarity between hand-crafted morphometrics and deep features is underutilized.

Purpose of the Study:

• To develop a multi-level feature fusion network for improved neuronal morphology description and classification.
• To leverage the complementarity of diverse feature representations.
• To enhance the distinctiveness of neuronal morphology descriptors.

Main Methods:

• Proposed a Multi-Level Fusion Module (MLFM) integrated into feature extraction blocks.
• MLFM includes a Feature Enhancement Module (FEM) using channel attention.
• MLFM incorporates a Feature Interaction Module (FIM) using cross-attention for feature complementarity.
• Developed a multi-level feature fusion network for neuronal morphology classification.

Main Results:

• Achieved 95.18% accuracy in classifying 10-type neurons on the NeuronMorpho-10 dataset.
• Outperformed existing single-feature or simple concatenation methods.
• Demonstrated strong performance and good generalization on NeuronMorpho-12 and NeuronMorpho-17 datasets.

Conclusions:

• The proposed multi-level feature fusion network effectively utilizes feature complementarity for superior neuronal morphology description.
• The method significantly improves the accuracy and generalization of neuron classification.
• This approach offers a more distinctive descriptor for characterizing neurons.