A deep learning algorithm for fast and accurate, non-rigid image registration that does not require a training data set.

Medical image registration is necessary for aligning multiple clinical images from a single patient as well as for associating medical images from multiple subjects or patients in research studies. Conventional medical image registration can be time consuming and challenging for certain types of images. Therefore, new methods use deep learning techniques like convolutional neural networks, which often employ training image datasets to learn informative image features. These techniques can improve the speed and accuracy of registration. However, in some cases training datasets are unavailable or do not accurately represent the image at hand.

To overcome these challenges, this image registration algorithm uses deep learning methods to identify the appropriate spatial transformations required to align medical images. While other deep learning methods have relied on previously acquired training data, this method directly trains fully convolutional networks to independently estimate spatial transformations.

To account for potentially large deformations between images, a multi-resolution strategy is adopted to jointly learn spatial transformations at different spatial resolutions. Because the registration of pairs of images also serves as a training procedure, the algorithm simultaneously optimizes and learns spatial transformations for the image registration.

• Fast image registration (~200 ms per image pair)
• Improved image alignment compared to current state-of-the-art techniques
• Multi-resolution registration strategy that accounts for deformations at different scales

• Medically necessary imaging that requires fast image registration
• Large clinical research studies that require accurate image registration