Quick start¶
These guides use pretrained weights that are downloaded on first run. All you need is a CUDA-enabled PyTorch and an editable install of MIR (Installation).
Affine pre-alignment with AffineReg3D, then deformable registration with VFA,
TransMorph-TVF or ConvexAdam-MIND, and flow composition.
Evaluate seven registration models from one notebook and compare Dice and folding.
Construct a population brain template from LUMIR24 by iterated registration and averaging.
The five-line pattern¶
Every registration model in MIR follows the same pattern. Once you know it, every tutorial reads the same way.
from MIR.models import TransMorphTVF, SpatialTransformer
import MIR.models.configs_TransMorph as configs_TransMorph
config = configs_TransMorph.get_3DTransMorph3Lvl_config() # 1. pick a configuration
config.img_size = (80, 112, 96) # 2. tell it the grid it will see
model = TransMorphTVF(config, time_steps=7).cuda() # 3. build the model
flow = model((moving, fixed)) # 4. predict a displacement in voxels
warped = SpatialTransformer(config.img_size).cuda()(moving, flow) # 5. warp anything with it
The Conventions page explains tensor shapes, displacement units, resolution changes and flow composition in detail.
Where to go next¶
Train your own model: Training a model walks through a complete training loop, and Tutorials lists the ready-made training scripts per dataset.
Pick a model: Models compares the architectures and their variants.
Look up a function: API reference.