DeepNAT_Feb.2017

DeepNAT: Deep Convolutional Neural Network

• 3-D patch based.
• Predict the central and the neighbors voxel of the patch.(Multi-task approach)
• 2 Netwroks hierarchically: separates foreground from background; identify 25 brain structures on the foreground.
• Indroduce intrinsic parameterization of the brain volume, formed by eigenfunctions of the Laplace-Beltrami operator to define the spatial context.

Network Architecture

• 3 convolutional layers with pooling, bathch normalization, and non-linearities, followed by fully connected layers with dropout.
• 2.7 million parameters