Deep Forest With Multi-Channel Message Passing and Neighborhood Aggregation Mechanisms for Brain Network...

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Deep learning methods like deep forest show promise for classifying brain networks (BNs) using functional magnetic resonance imaging (fMRI) data, especially with limited sample sizes.
• Existing deep forest approaches for BNs often use sliding windows, neglecting prior knowledge and failing to capture complex topological features crucial for disease-related alterations.

Purpose of the Study:

• To propose a novel deep forest framework (DF-MCMPNA) that effectively extracts and aggregates long-range, multi-channel topological features from brain networks for improved classification.
• To leverage intrinsic connectivity networks (ICNs) and whole-brain data to enhance feature representation for identifying brain disorders.

Main Methods:

• Developed a deep forest framework (DF-MCMPNA) incorporating multi-channel message passing and neighborhood aggregation mechanisms.
• Utilized four feature extraction channels: three intrinsic connectivity networks (ICNs) and the whole-brain network.
• Employed a multi-channel message passing mechanism for local topological feature learning and a channel-shared neighborhood aggregation mechanism for fusing multi-channel features.

Main Results:

• The proposed DF-MCMPNA framework demonstrated superior classification performance compared to state-of-the-art methods on the ABIDE I, ABIDE II, and ADHD-200 datasets.
• The method successfully identified abnormal brain regions, indicating its potential for clinical applications.
• DF-MCMPNA effectively extracts and aggregates long-range multi-channel topological features, outperforming traditional sliding window approaches.

Conclusions:

• The DF-MCMPNA framework offers a significant advancement in classifying brain networks using fMRI data, particularly in scenarios with limited samples.
• The integration of multi-channel message passing and neighborhood aggregation enhances the capture of complex topological brain network features relevant to neurological disorders.
• This approach holds promise for improving the accuracy and efficiency of diagnosing brain diseases through neuroimaging analysis.