Modeling Functional Brain Networks for ADHD via Spatial Preservation-Based Neural Architecture Search

Area of Science:

• Neuroimaging and Computational Neuroscience
• Medical Image Analysis
• Machine Learning for Healthcare

Background:

• Abnormal functional connectivity in functional brain networks (FBNs) is observed in attention deficit hyperactivity disorder (ADHD) compared to typical controls (TC).
• Existing FBN modeling methods using dimensionality reduction on fMRI data can lead to misdiagnosis due to loss of spatial information and voxel confusion.

Purpose of the Study:

• To propose a novel spatial preservation-based neural architecture search (SP-NAS) for more accurate FBN modeling in ADHD.
• To address limitations of dimensionality reduction in fMRI data processing for ADHD diagnosis.
• To explore cross-regional association differences between ADHD and TC groups for auxiliary diagnosis.

Main Methods:

• Developed a spatial preservation module to embed original spatial information into dimensionality reduction data, mitigating voxel confusion.
• Constructed a search space with optimized operations for efficient extraction of spatial-temporal fMRI data characteristics in ADHD.
• Utilized the ADHD-200 dataset for model validation and comparison with typical controls.

Main Results:

• The proposed SP-NAS model achieved competitive results in ADHD diagnosis.
• The model successfully identified abnormal connections in lesion regions of ADHD, consistent with clinical findings.
• Preservation of spatial information improved the accuracy of FBN modeling, reducing misdiagnosis risks.

Conclusions:

• SP-NAS offers a promising approach for accurate FBN modeling in ADHD diagnosis by preserving crucial spatial information.
• The method enhances diagnostic accuracy and provides insights into the neurobiological underpinnings of ADHD.
• This technique can aid in the auxiliary diagnosis of ADHD by identifying specific abnormal brain network connections.