Learning and combining image neighborhoods using random forests for neonatal brain disease classification

Area of Science:

• Neuroimaging
• Developmental Neuroscience
• Machine Learning

Background:

• Characterizing early childhood brain development is complex due to heterogeneous data.
• Manifold learning techniques reduce data complexity by finding low-dimensional representations.
• Neighborhood definition is critical for manifold learning but application-dependent.

Purpose of the Study:

• To develop a framework for learning and optimally combining multiple pairwise distances from brain image data.
• To create a representation that preserves multiple distances from heterogeneous sources.
• To identify predictive features associated with these distances for improved classification.

Main Methods:

• Proposed a framework to learn multiple pairwise distances in brain image populations.
• Combined these distances in an unsupervised manner within a manifold learning step.
• Utilized neighborhood approximation forests for neighborhood structure learning.
• Selected predictive features associated with the learned distances.

Main Results:

• The proposed method effectively combines multiple distances from heterogeneous sources.
• The combined distances yield a population representation that preserves multiple similarity structures.
• The method successfully selects predictive features related to the distances.
• In neonatal magnetic resonance images, combining multiple distances improved characterization and classification of clinical groups compared to single distances.

Conclusions:

• Combining multiple, heterogeneous distance measures enhances manifold learning for brain development studies.
• The framework offers a robust approach for analyzing complex neuroimaging data.
• This method improves the classification accuracy of early childhood brain conditions.