Medical image registration involves overlaying two images to compare and analyse differences – such as changes in a tumour over time – in great detail. The process, however, can often take two hours or more using traditional systems. In a pair of upcoming conference papers, researchers from MIT describe a machine-learning-based algorithm that can register brain MR scans and other 3D images more than 1000 times faster.
While existing algorithms start from scratch for every pair of images, the new algorithm, called VoxelMorph, speeds the process up by “learning” as it registers image pairs. In doing so, it acquires information about how to align images and estimates some optimal alignment parameters. After training, the algorithm uses those parameters to map all pixels of one image to another at once.
“The tasks of aligning a brain MRI shouldn’t be that different when you’re aligning one pair of brain MRIs or another,” says Guha Balakrishnan, a graduate student at MIT. “There is information you should be able to carry over in how you do the alignment. If you’re able to learn something from previous image registration, you can do a new task much faster and with the same accuracy.”
In a paper presented today at the Conference on Computer Vision and Pattern Recognition, the researchers describe how they trained their algorithm on 7000 MRI brain scans and then tested it on 250 additional scans.
During training, pairs of brain scans were fed into the algorithm, which captured similarities of voxels in the two scans. In doing so, it learns information about groups of voxels – such as anatomical shapes common to both scans – which it uses to calculate optimized parameters. When fed two new scans, the algorithm uses the optimized parameters to rapidly calculate the exact alignment of every voxel in both scans.
The researchers found that their algorithm accurately registered all 250 test brain scans within two minutes using a traditional central processing unit, and in under one second using a graphics processing unit. They note that, importantly, the algorithm is “unsupervised”, meaning that it doesn’t require additional information such as ground truth data or anatomical landmarks.
The second paper, to be presented at MICCAI in September, will describe a refined VoxelMorph algorithm that validates the accuracy of each registration. It also guarantees the registration “smoothness”, so that it doesn’t produce folds, holes or general distortions in the composite image. Across 17 brain regions, the refined algorithm scored the same accuracy as a state-of-the-art 3D registration algorithm, while providing runtime and methodological improvements.
The algorithm has a wide range of potential applications, the team points out. MIT colleagues, for instance, are currently running the algorithm on lung images. It could also pave the way for image registration during operations, potentially enabling surgeons to register scans in near real-time.