NON-EUCLIDEAN, CONVOLUTIONAL LEARNING ON CORTICAL BRAIN SURFACES.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Cortical features from surface vertices show promise for detecting neurodevelopmental and neurodegenerative diseases.
• High-dimensional surface data presents challenges for training stable classifiers due to limited datasets.
• Current feature reduction methods using standard brain atlases may lose critical pathological information.

Purpose of the Study:

• To propose a novel data-driven approach for feature reduction and improved classification on cortical surfaces.
• To extend convolutional neural networks (CNNs) for application on non-Euclidean manifolds like cortical surfaces.
• To develop a more effective method for classifying neurodegenerative diseases using brain surface data.

Main Methods:

• Developed a surface-CNN framework to learn powerful features and relevant brain regions from high-dimensional cortical data.
• Implemented a data-driven approach to reduce feature dimensionality and enhance feature separation.
• Applied the framework to classify Alzheimer's disease patients versus normal controls.

Main Results:

• The proposed surface-CNN framework successfully reduced dimensionality and improved feature separation.
• Achieved high performance in cross-validation for classifying Alzheimer's disease.
• Demonstrated the potential of the novel prediction system for disease diagnosis.

Conclusions:

• The surface-CNN framework offers a promising approach for analyzing complex brain surface data.
• This method effectively addresses the challenges of high-dimensional data in neuroimaging.
• The framework shows significant potential for accurate disease classification in clinical applications.