Communications of the ACM
Faster Analysis of Medical Images
Massachusetts Institute of Technology researchers have developed a machine-learning algorithm capable of registering brain scans and other three-dimensional images about 1,000 times faster than previous methods through the use of novel learning methods.
Credit: Guha Balakrishnan et al.
Massachusetts Institute of Technology researchers have developed a machine-learning algorithm capable of registering brain scans and other three-dimensional images about 1,000 times faster through the use of novel learning methods.
The VoxelMorph algorithm "learns" while registering thousands of pairs of images, a process through which it obtains data about how to align images and calculates optimal alignment parameters. Once training is completed, it uses those parameters to map the pixels of one image to another all at once, slashing registration time to one or two minutes on a normal computer, or under one second on a graphics processing unit with comparable accuracy to cutting-edge systems.
VoxelMorph is driven by a convolutional neural network comprised of multiple nodes that process images and other data across several computational layers. The unsupervised algorithm requires no additional information beyond image data.
From MIT News
View Full Article
Abstracts Copyright © 2018 Information Inc., Bethesda, Maryland, USA
No entries found