Doctors often need to compare two MRI images to track changes in the body over time, but the process of lining up the images to make accurate measurements is extremely time-consuming. It can take hours for a computer to match all the locations in a 3D map, but researchers from MIT have developed an algorithm that could cut that time to less than a second.
MRI scans are cumbersome to manage because of how much information they contain. Each scan is essentially hundreds of stacked 2D images. These form the 3D map known as a volume. The volume is made up of 3D pixels known as voxels. When a computer aligns two different MRI scans, it’s sifting through millions of voxels to assign them locations in a new, unified image. Scans can also come from different machines with varying spatial properties, slowing the work even more.
Several hours of computing time is considered quite good for MRI analysis. Researchers trying to analyze data from large populations across multiple patients with the same disease can end up waiting hundreds of hours for a computer to generate aligned images. Simply throwing more processing power at the problem isn’t practical, but the “VoxelMorph” system from MIT researchers might do the trick.
VoxelMorph is a convolutional neural network, so the team started by training it with 7,000 publicly available MRI brain scans. In a neural network, you add data at one end, and the network passes it through numerous nodes that feed forward into other nodes. Depending on the weighting of each node, you end up with an output that should provide the desired results. VoxelMorph learned about common groups of voxels and anatomical shapes.
After training, the team used 250 new scans to test the network’s effectiveness. VoxelMorph completed in two minutes what would have taken a conventional MRI analysis program several hours to do. That’s just with a regular CPU. When VoxelMorph runs on a GPU, the process takes less than a second. If you were building a machine to process MRI images, you’d probably configure it to run calculations on the GPU.
So, we’ve effectively gone from hours to instantaneous. The team suggests this could change the way doctors perform some surgeries. It may be possible to make new scans during a surgery and get real-time analysis of the images. The system could also potentially work on other types of 3D scans with additional training.
Google’s AI-Focused Tensor Processing Units Now Available in Beta
Google is ready to open up its Cloud TPU platform to developers and researchers looking to test machine learning workloads — and it's got a new, more powerful Cloud TPU design than the chips we've previously discussed.
ARM Announces Project Trillium, a New Dedicated AI Processing Family
ARM has launched a new series of AI-focused processors, with the goal of improving machine learning in mobile devices.
MediaTek Unveils New P60 Helio With Integrated AI Processing
MediaTek is unveiling its new Helio P60 SoC at Mobile World Congress, with substantially improved CPU, GPU, and AI performance.