BNCPL: Brain-Network-based Convolutional Prototype Learning for Discriminating Depressive Disorders.

Area of Science:

• Neuroimaging
• Machine Learning
• Psychiatry

Background:

• Deep learning excels at identifying patterns in complex neuroimaging data.
• Distinguishing depressive disorders from healthy controls using neuroimaging remains challenging.
• Existing methods often struggle with subtle group differences in high-dimensional data.

Purpose of the Study:

• To develop a novel deep learning model, the brain-network-based convolutional prototype learning (BNCPL) model.
• To enhance the classification performance for depressive disorders using resting-state functional connectivity (FC).
• To improve model interpretability by identifying discriminative brain circuits.

Main Methods:

• Developed the BNCPL model, integrating convolutional neural networks and prototype learning.
• Applied BNCPL to resting-state functional connectivity data from 208 depressive disorder patients and 210 healthy controls across multiple sites.
• Utilized saliency maps to visualize and identify discriminative functional connections.

Main Results:

• Achieved 71.0% accuracy in multi-site classification of depressive disorders versus healthy controls.
• Demonstrated a 2.4-7.2% accuracy improvement over traditional classifiers and alternative deep learning models.
• Identified prefrontal-subcortical circuits as key discriminators, correlating with disease severity and cognitive ability.

Conclusions:

• The BNCPL model offers improved interpretability and classification performance for neuroimaging data.
• Dysregulation in functional prefrontal-subcortical circuits is a significant factor in discriminating depressive disorders.
• Integrating prototype learning with saliency maps advances deep learning applications in psychiatric research.