Sparse Deep Neural Network for Encoding and Decoding the Structural Connectome.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Deep learning for neuroimaging faces challenges with high-dimensional data and limited samples.
• Accurate classification of brain states is crucial for neurodegenerative disease diagnosis.

Purpose of the Study:

• To develop a sparse deep neural architecture for efficient encoding and decoding of human brain structural connectomes.
• To improve classification accuracy for Alzheimer's disease (AD) and Parkinson's disease (PD) using DTI brain scans.
• To identify key biomarkers associated with AD and PD through advanced feature selection methods.

Main Methods:

• A novel sparse feedforward deep neural architecture with a sparsely connected element-wise multiplication first hidden layer and a fixed transform output layer.
• Application of DeepLIFT, Layer-wise Relevance Propagation (LRP), and Integrated Gradients (IG) for feature relevance analysis.
• Recursive Feature Elimination (RFE) algorithm for identifying key biomarkers and removing irrelevant features.

Main Results:

• The proposed sparse architecture significantly reduced trainable parameters (45.1% for AD, 47.1% for PD) and training time compared to standard feedforward networks.
• Classification accuracy increased by 2.6% (AD) and 3.1% (PD) for cognitively normal (CN) vs. disease classification.
• The RFE method further improved accuracy (2.1% for AD, 4% for PD) while removing 90-95% of irrelevant features, identifying biomarkers consistent with existing literature.

Conclusions:

• The sparse deep neural architecture offers a computationally efficient and accurate method for brain connectome analysis.
• Relevancy score-based methods are effective for brain decoding and identifying disease-specific biomarkers.
• The approach successfully reduces model complexity, enhances classification performance, and detects biologically relevant brain regions and connections.