Deep Manifold Harmonic Network With Dual Attention for Brain Disorder Classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Accurate analysis of neurological disorders aids early diagnosis of brain and psychiatric conditions, often linked to brain atrophy.
• Geometric deep learning offers novel ways to analyze brain network geometry, but faces challenges with data heterogeneity and noise.
• Identifying discriminative features is crucial for improving diagnostic accuracy in complex brain network data.

Purpose of the Study:

• To present a novel dual-attention deep manifold harmonic discrimination (DA-DMHD) method for the early diagnosis of neurodegenerative diseases.
• To address the limitations of existing geometric deep learning methods in handling heterogeneous and noisy brain network data.
• To enhance the identification of discriminative features for improved diagnostic accuracy.

Main Methods:

• Learning a low-dimensional manifold projection to exploit geometric features of brain networks.
• Developing attention blocks with discrimination to learn representations that facilitate group-dependent discriminant matrices.
• Utilizing group-specific references for downstream analysis.

Main Results:

• The DA-DMHD model was evaluated on the ADNI and ADHD-200 datasets.
• Experimental results show the model effectively handles heterogeneous brain network topological differences.
• The proposed method achieved excellent classifying performance in both accuracy and robustness compared to state-of-the-art methods.

Conclusions:

• The DA-DMHD method offers a robust approach for early diagnosis of neurodegenerative diseases.
• The model's ability to capture complex topological differences in brain networks signifies a significant advancement.
• This work contributes to improving diagnostic accuracy and robustness in neurodegenerative disease detection.