Anatomy-aware and acquisition-agnostic joint registration with SynthMorph.

Area of Science:

• Medical image analysis
• Computational neuroscience
• Machine learning in healthcare

Background:

• Classical affine registration is accurate but slow.
• Deep learning (DL) methods are fast but struggle with large transforms and domain shifts.
• Current methods often lack anatomy-specificity and require preprocessing.

Purpose of the Study:

• Introduce SynthMorph, a DL tool for fast, joint affine-deformable brain image registration.
• Address limitations of existing registration methods, including speed, accuracy, and anatomy-awareness.
• Provide an end-to-end solution for brain MRI registration without preprocessing.

Main Methods:

• Utilize a synthesis strategy training networks with label maps for robustness across unseen acquisitions.
• Optimize spatial overlap of anatomical labels for anatomy-specific registration.
• Combine affine models with a deformable hypernetwork for user-defined deformation-field regularity.

Main Results:

• SynthMorph demonstrates high accuracy in affine and deformable registration.
• Achieves robust performance across diverse neuroimaging data and acquisition specifics.
• Offers registration in a fraction of the time required by classical methods.

Conclusions:

• SynthMorph provides a fast, symmetric, diffeomorphic, and user-friendly DL solution for brain image registration.
• The framework enables anatomy-aware, acquisition-agnostic registration for various anatomies and architectures.
• SynthMorph is a complete end-to-end solution for brain MRI registration.