Sparse non-convex regularization based explainable DBN in the analysis of brain abnormalities in schizophrenia

Area of Science:

• Medical Image Analysis
• Neuroscience
• Machine Learning

Background:

• Deep belief networks (DBNs) are common in medical imaging but struggle with high-dimensional, small-sample data, leading to overfitting.
• Traditional DBNs prioritize performance over explainability, a critical factor in medical applications.
• Existing models face challenges with dimensional disaster and lack interpretability.

Purpose of the Study:

• To develop a sparse, non-convex, explainable deep belief network (DBN) for enhanced medical image analysis.
• To address overfitting and improve generalization in DBNs using non-convex sparsity learning.
• To integrate feature selection for improved decision-making and interpretability in medical data.

Main Methods:

• Proposed a novel sparse non-convex deep belief network by integrating non-convex regularization and Kullback-Leibler divergence penalty.
• Implemented feature back-selection based on layer weight row norms for identifying crucial decision-making features.
• Applied the model to schizophrenia brain imaging data for performance evaluation and feature identification.

Main Results:

• The proposed explainable DBN demonstrated superior performance compared to traditional feature selection models.
• The model effectively reduced model complexity and enhanced generalization ability.
• Identified 28 functional brain connections significantly correlated with schizophrenia.

Conclusions:

• The developed sparse non-convex explainable DBN offers improved performance and generalization for medical image analysis.
• Feature back-selection successfully identified critical biomarkers for schizophrenia, aiding diagnosis and treatment.
• This approach provides a robust methodological foundation for analyzing similar brain disorders.