Combining medical images of patients can provide interesting insights, but is a very time-consuming process. A new technique based on a self-learning algorithm is helping to speed up this process considerably.
Image registration is a widely used technique that combines several medical images. For example, by superimposing several MRI scans taken at different times, it is possible, among other things, to visualise in detail small changes in the development of a tumour. It is also possible to compare medical images of different patients with the same disease to discover patterns in its development.
Combining millions of 3D pixels
Combining medical images is a time-consuming process, which limits the applicability of the technique. For instance, an MRI scan actually consists of hundreds of stacked 2D images that combine to form a large 3D image. This image is also known as a 'volume'. A volume easily contains a million or more 3D pixels, called 'voxels'. To compare medical images properly, it is necessary to accurately align the voxels of one volume with the voxels of the other volume.
Aligning voxels is a complex and time-consuming process. Not only may medical images have been taken by different medical devices, the orientation of medical images may also differ. Such differences have to be compensated by the image registration system, which requires a lot of computing power. Consequently, image registration systems quickly need several hours to combine medical images.
In particular, comparing medical images of large groups of patients poses problems. For example, comparing medical images of hundreds of patients quickly takes several hundred hours. According to the Massachusetts Institute of Technology (MIT), this is mainly due to the fact that algorithms using traditional image capture systems are unable to learn. "In essence, they start completely from scratch for each new pair of medical images," says Guhu Balakrishan, PhD, affiliated with the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Department of Engineering and Computer Science (EECS) at MIT.
Optimal values
A new machine learning algorithm from MIT is capable of learning, though, and can significantly speed up image capture as a result. Called 'VoxelMorph', the algorithm was trained by researchers by having it combine thousands of pairs of medical images of brains. A total of 7,000 medical images were registered during the training phase. During this process, the algorithm gathered information on how the images were combined with each other and, based on this, estimates the optimal values that can be used for this purpose.
Researchers had VoxelMorph combine 250 medical images of brains during a test. This showed that, thanks to its training, the algorithm is able to capture medical images in about two minutes on a regular computer system without a graphics card. If a graphics card is deployed, the process can even be performed in less than a second. This while the accuracy of VoxelMorph is not inferior to that of traditional image registration systems.
Also suitable for other medical images
VoxelMorph has been trained to combine with medical images of brains. However, the researchers argue that the algorithm can also be trained for combining other medical images. Within MIT, for instance, it is already experimenting with registering images of lungs.
The researchers argue that their algorithm offers opportunities in the operating theatre, among others. For instance, surgeons could record scans more or less in real-time to get a much clearer picture of the progress of an operation. "Currently, they cannot compare images during an operation, as this takes two hours and the operation has to continue," said Adrian Dalca, a postdoctoral fellow at Massachusetts General Hospital and MIT's CSAIL. "If this only takes a second, you can imagine that this is feasible."
Author: Wouter Hoeffnagel
Source: MIT (photo)
Source: Paper 'An Unsupervised Learning Model for Deformable Medical Image Registration'
