Rotation-invariant graph message passing enables acquisition protocol generalisation in learning-based brain...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Estimating brain microstructure is crucial for medicine and neuroscience.
• Diffusion-weighted magnetic resonance imaging (DW-MRI) allows in vivo measurement.
• Current methods like biophysical model fitting are slow and impractical for clinical settings.

Purpose of the Study:

• To develop a rapid and versatile method for brain microstructure estimation.
• To overcome limitations of existing machine learning methods that require retraining for new MRI acquisition protocols.

Main Methods:

• A graph neural network (GNN) was developed, treating DW-MRI data as a 3D point cloud.
• The GNN incorporates rotation-invariant message passing and permutation-invariant pooling.
• Inductive biases were guided by physics and symmetries, rather than generic architectures.

Main Results:

• The GNN demonstrated domain generalization, accurately estimating microstructure from unseen real-world protocols.
• No retraining was necessary when encountering new acquisition protocols.
• The model produces fixed-size embeddings encoding microstructure information.

Conclusions:

• The proposed GNN offers a "train once, deploy anywhere" solution for microstructure estimation.
• This approach accelerates the clinical deployment of microstructure mapping using machine learning.
• The method is robust to arbitrary acquisition protocols, enhancing practical utility.