SynthMorph

SynthMorph performs contrast-agnostic MRI image registration by training networks on synthetic shapes, label maps, and images generated from noise distributions to enable robust 3D neuroimaging alignment without relying on acquired intensity images.

Key Features:

• Contrast-Agnostic Registration: Trains networks to register MRI images irrespective of imaging contrast, enabling cross-contrast alignment.
• Generative Synthetic Data: Synthesizes diverse label maps and images from arbitrary noise distributions to expose networks to extensive variability during training.
• Independence from Acquired Data: Eliminates reliance on acquired intensity images by using synthetic shapes and images for network training.
• Use of Anatomical Label Maps: Incorporates synthesized anatomical label maps when available to improve registration accuracy without real intensity images.
• Single-Model Cross-Contrast Performance: Achieves registration accuracy across and within contrasts using a single trained model, surpassing prior state-of-the-art in reported experiments.

Scientific Applications:

• 3D Neuroimaging Registration: Aligns 3D MRI neuroimaging data across arbitrary contrasts for structural and morphometric analyses.
• Registration without Intensity Images: Enables registration workflows when only label maps or synthetic representations are available, supporting studies lacking acquired intensity contrasts.

Methodology:

Networks are trained using synthetic data generated from noise distributions, including synthesized anatomical label maps when available, to expose models to wide variability and enforce contrast-agnostic registration.