Topology-guided cyclic brain connectivity generation using geometric deep learning.

Area of Science:

• Neuroscience
• Computer Science
• Medical Imaging

Background:

• Analyzing medical data like brain connectomes is crucial but limited by small sample sizes, risking model overfitting.
• Existing deep learning (DL) methods for data synthesis are often designed for Euclidean data, neglecting the geometric nature of brain connectomes.
• Current geometric DL approaches for brain connectome prediction focus on domain alignment, ignoring the preservation of connectome topology.

Purpose of the Study:

• To develop a novel deep learning framework for brain connectome synthesis that addresses limitations of existing methods.
• To enable bidirectional translation between different views of brain networks while preserving topological properties.
• To improve data augmentation strategies for clinical diagnosis using synthetic medical data.

Main Methods:

• Adapted the graph translation generative adversarial network (GT GAN) for brain connectomic data.
• Extended GT GAN to a cyclic graph translation (CGT) GAN for bidirectional network translation.
• Introduced a topological strength constraint to create the CGTS GAN, preserving the topology of brain regions of interest (ROIs).

Main Results:

• The proposed CGTS GAN architecture was compared against baseline graph translation methods and ablated versions.
• Our deep graph network demonstrated superior performance, achieving lower mean squared error (MSE) on a multiview autism spectrum disorder connectomic dataset.
• The results indicate improved accuracy in synthesizing brain connectomes compared to existing methods.

Conclusions:

• A topology-aware, bidirectional brain connectome synthesis framework was designed using geometric deep learning.
• The CGTS GAN effectively synthesizes brain connectomes while preserving essential topological features.
• This framework offers a valuable tool for data augmentation in clinical diagnosis and medical data analysis.