voxelmorph

Why we included this project

Aligning medical images has traditionally meant slow iterative optimization, but VoxelMorph replaces that with a deep network that learns a deformation field warping one volume onto another. Its unsupervised losses remove the need for manually labeled training data, and diffeomorphic transforms keep the resulting warps invertible and biologically plausible. The library ships training, registration, and evaluation scripts, including Dice overlap scoring, plus pretrained models, so a team can register a pair of volumes quickly or retrain on its own data with modest customization. It is aimed at medical imaging researchers and engineers building pipelines around longitudinal or population scans, and the tutorial plus several Colab demos, deformable and affine registration among them, make it approachable. One caveat: the PyTorch version is under active development and interfaces may shift, so teams wanting stability may prefer the TensorFlow branch.