SynthMorph: Learning Contrast-Invariant Registration Without Acquired Images.

Area of Science:

• Medical imaging
• Artificial intelligence
• Computational anatomy

Background:

• Classical image registration is computationally intensive, requiring optimization for each new image pair.
• Existing learning-based methods are limited by the contrast and content of training data.
• Magnetic resonance imaging (MRI) poses challenges due to inherent contrast variability.

Purpose of the Study:

• To develop a data-agnostic learning strategy for image registration.
• To create robust neural networks capable of generalizing across diverse MRI contrasts.
• To overcome limitations of traditional and current learning-based registration techniques.

Main Methods:

• Leveraging generative models to create diverse synthetic label maps and images for training.
• Synthesizing training data from noise distributions and anatomical label maps.
• Training deep learning networks on this synthetic data to learn invariant features.

Main Results:

• Achieved state-of-the-art registration accuracy across various MRI contrasts using a single model.
• Demonstrated robust generalization to unseen contrasts, even those not present during training.
• Showcased competitive performance using synthesized data from noise, and significantly boosted performance using anatomical label maps.

Conclusions:

• The proposed generative strategy enables powerful, contrast-agnostic image registration networks.
• This approach eliminates the need for acquired imaging data during training, enhancing flexibility.
• The method offers a significant advancement for medical image analysis and neuroimaging applications.