AutoAtlas: Neural Network for 3D Unsupervised Partitioning and Representation Learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Accurate partitioning and representation of 3D brain MRI volumes are crucial for understanding brain structure and function.
• Existing methods often require supervision or yield high-dimensional, less interpretable features.

Purpose of the Study:

• To introduce AutoAtlas, a novel neural network architecture for fully unsupervised partitioning and representation learning of 3D brain MRI volumes.
• To enable subject-specific brain tissue analysis and metadata prediction using learned features.

Main Methods:

• Developed a dual-component neural network (AutoAtlas) for simultaneous multi-label partitioning and information compression.
• Trained the network using a loss function promoting accurate reconstruction, spatial smoothness, and contiguous partitions.
• Extracted low-dimensional features representing local texture within each partition.

Main Results:

• AutoAtlas successfully partitions 3D brain MRI volumes, adapting to subject-specific structural variations while maintaining consistent spatial locations across subjects.
• The derived features are low-dimensional and effectively represent local tissue texture.
• Metadata prediction using AutoAtlas features showed comparable or superior performance to FreeSurfer-derived features.
• Partition-specific feature importance scores can be visualized on the brain.

Conclusions:

• AutoAtlas offers a powerful unsupervised approach for brain MRI analysis, providing both accurate partitioning and meaningful feature representation.
• The method facilitates subject-specific analysis and enables novel visualization of feature importance within brain partitions.
• AutoAtlas has the potential to advance neuroimaging research by offering an efficient and interpretable tool for exploring complex brain data.