Topology-Guided Graph Masked Autoencoder Learning for Population-Based Neurodevelopmental Disorder Diagnosis

Area of Science:

• Neuroscience
• Computational Psychiatry
• Medical Imaging Analysis

Background:

• Understanding brain disorder mechanisms requires analyzing population-level data.
• Current methods often overlook individual neural circuit information and inter-individual associations.
• Existing approaches may use clinical data or data augmentation but lack focus on individual representation.

Purpose of the Study:

• To propose a novel method, Topology-guided Graph Masked autoencoder Learning (TGML), for detecting abnormal neural circuits in brain diseases.
• To focus on individual representation and intra-population associations for improved diagnosis.
• To enhance the diagnosis of neurodevelopmental disorders within a population.

Main Methods:

• Developed the Topology-guided Graph Masked autoencoder Learning (TGML) method.
• Implemented a topology-guided group association module (T${G}^{{2}}$AM) for graph reconstruction and edge updates.
• Utilized an intra-population interaction masked autoencoder network (IPI_MAE) for capturing subject characteristics.

Main Results:

• TGML demonstrated significant improvements in diagnostic accuracy for neurodevelopmental disorders.
• The method effectively captures discriminative characteristics of subjects.
• TGML surpassed existing state-of-the-art methods in Autism Spectrum Disorder and Attention Deficit Hyperactivity Disorder diagnosis tasks.

Conclusions:

• The proposed TGML method offers a novel and effective approach for diagnosing brain diseases by focusing on individual and population-level neural circuit analysis.
• TGML advances the field of computational neuroscience by integrating graph learning and masked autoencoders for disease detection.
• This approach holds promise for improving the diagnosis and understanding of neurodevelopmental disorders.