Bundle geodesic convolutional neural network for diffusion-weighted imaging segmentation.

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Neuroscience

Background:

• Machine learning on diffusion-weighted imaging (DWI) data is hindered by large sample sizes and limited labeled data.
• Leveraging the inherent geometry of DWI data allows for learning patterns from minimal training sets.

Purpose of the Study:

• To develop a novel Riemannian deep learning framework for tissue classification using single-shell DWI data.
• To address the challenges of data size and scarcity of labels in DWI analysis.

Main Methods:

• A three-layer framework: lifting layer for local representation on tangent spaces, group convolution layer for rotation kernel convolution, and projection layer for manifold-based function formation.
• The approach utilizes the geometry of DWI data within a Riemannian deep learning architecture.

Main Results:

• The proposed method achieves state-of-the-art performance with significantly fewer model parameters.
• Sensitivity analysis demonstrated that reduced training data (down to 29.4%) mildly impacts overall accuracy but improves minority class performance.

Conclusions:

• This work extends convolutional neural networks to Riemannian manifolds, offering a powerful tool for brain structural pattern analysis.
• The framework shows potential for aiding manual data annotation in neuroimaging.