A fully inductive inference protocol for population GNNs in single-subject brain disorder diagnosis

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Population graph-based Graph Neural Networks (GNNs) excel in brain disease diagnosis by modeling inter-subject relationships.
• Existing transductive GNNs perform poorly on unseen subjects, while inductive models struggle with single-subject inference.
• Current methods often require test batches or lack generalization for individual unseen nodes.

Purpose of the Study:

• To propose a fully inductive inference protocol for population graphs specifically for single-subject brain disease diagnosis.
• To overcome the limitations of transductive and existing inductive GNNs in handling unseen subjects and enabling efficient individual inference.
• To develop a method that avoids the retraining bottleneck associated with transductive models.

Main Methods:

• Constructing a population graph solely with training nodes.
• Dynamically connecting a single unseen test subject to the training graph during inference using imaging and phenotypic similarities.
• Evaluating the protocol on multiple neuroimaging datasets (ABIDE I, ABIDE II, ADHD-200).

Main Results:

• The proposed fully inductive method outperforms state-of-the-art transductive and inductive baselines in a rigorous inductive evaluation.
• Single-subject inference maximizes diagnostic performance by minimizing interference between test subjects.
• The approach eliminates the need for retraining, offering practical advantages for real-time deployment.

Conclusions:

• The developed inductive inference protocol offers a robust and efficient solution for single-subject brain disease diagnosis using population graphs.
• This method provides a significant operational advantage for clinical deployment and real-time inference workflows.
• The findings highlight the potential of fully inductive GNNs for advancing neuroimaging-based diagnostics.