voxelmorph/voxelmorph

What it solves

VoxelMorph provides a general-purpose library for image registration and alignment, specifically focusing on learning-based tools to model deformations. It addresses the challenge of aligning different images (such as medical scans) by learning how to warp one image to match another, which is often more efficient than traditional iterative registration methods.

How it works

The library uses convolutional neural networks (CNNs) to predict deformation fields that map one image to another. It supports several registration paradigms:

• Unsupervised Learning: Training models using loss functions like Mean Squared Error (MSE) or Cross-Correlation (CC) without needing ground-truth labels.
• Diffeomorphic Registration: Integrating velocity fields to ensure the resulting deformations are smooth and invertible.
• SynthMorph: A strategy that allows training registration models using synthesized images (random shapes or brain label maps) rather than real acquired imaging data, making the models agnostic to image contrast.
• HyperMorph: A method to learn the effect of registration hyperparameters to avoid manual tuning.

Who it’s for

It is primarily designed for researchers and practitioners in medical imaging and computer vision who need to align 3D or N-dimensional images, create deformable templates (atlases), or perform multi-modal registration (e.g., CT to MRI).

Highlights

• Contrast-Agnostic Registration: Through SynthMorph, models can be trained without real data to handle various image contrasts.
• Flexible Framework: Supports both PyTorch and TensorFlow (via the dev-tensorflow branch).
• Template Construction: Includes tools for creating both unconditional and conditional deformable templates.
• Multi-format Support: Handles NIfTI, MGZ, and numpy (.npz) file formats.
• Pre-trained Models: Offers a variety of pre-trained weights for common registration tasks.