Decomposition-Based Correlation Learning for Multi-Modal MRI-Based Classification of Neuropsychiatric Disorders.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Multi-modal magnetic resonance imaging (MRI) is crucial for diagnosing brain diseases.
• High-dimensional MRI data presents challenges for deep learning models, especially when integrating multiple modalities.
• Existing methods struggle with the complexity of joint analysis of structural and functional MRI.

Purpose of the Study:

• To develop an effective method for analyzing complex relationships between structural MRI and functional MRI data.
• To improve the accuracy of classifying neuropsychiatric disorders using multi-modal MRI.
• To identify disease-specific feature connections within brain matrices.

Main Methods:

• Decomposition-based Correlation Learning (DCL) was developed, incorporating matrix decomposition principles.
• DCL considers spike magnitude of leading eigenvalues, sample size, and matrix dimensionality.
• Canonical Correlation Analysis (CCA) was employed for correlation analysis and matrix construction.

Main Results:

• DCL demonstrated higher classification accuracy for neuropsychiatric disorders compared to existing methods on the Consortium for Neuropsychiatric Phenomics (CNP) dataset.
• The method identified significant feature connections in brain matrices, differentiating between disease and normal cases, and disease subtypes.
• State-of-the-art performance was achieved across large and small sample size datasets.

Conclusions:

• Decomposition-based Correlation Learning (DCL) offers a robust approach for multi-modal MRI analysis in clinical practice.
• DCL effectively captures intricate relationships within MRI data, leading to improved diagnostic accuracy for brain diseases.
• The identified brain matrix features provide valuable insights for understanding and classifying neuropsychiatric disorders.