PROJECT TITLE :

AutoAtlas Neural Network for 3D Unsupervised Partitioning and Representation Learning

ABSTRACT:

We present a novel neural network architecture that we call AutoAtlas for fully unsupervised partitioning and representation learning of three-dimensional magnetic resonance imaging (MRI) volumes of the human brain. One neural network is used to perform multi-label partitioning based on the local texture in the volume, and another neural network is used to compress the information that is contained within each partition. The two neural network components that make up AutoAtlas are as follows: We train both of these components simultaneously by optimizing a loss function that is designed to promote accurate reconstruction of each partition, while also encouraging spatially smooth and contiguous partitioning, and discouraging relatively small partitions. All of this is accomplished by discouraging relatively small partitions and encouraging spatially smooth and contiguous partitioning. We show that the partitions are able to adapt to the subject-specific structural differences in the brain tissue while still appearing at the same spatial locations in all of the subjects' brains. In addition to this, AutoAtlas generates very low dimensional features that represent the local texture of each partition. We demonstrate the ability to predict metadata associated with each subject by using the derived feature representations, and we compare the results to the ability to predict using features derived from FreeSurfer's anatomical parcellation. Because our features are inherently linked to separate partitions, we are able to map values of interest onto the brain in order to visualize it. One example of this would be mapping partition-specific feature importance scores.