Models

MIR implements its own registration networks and wraps third-party methods behind the same interface (Conventions). Each model page lists the variants, a usage snippet, pretrained weights, the tutorials that train or evaluate it, and the paper to cite.

Registration networks

TransMorph

Swin-Transformer encoder with a convolutional decoder. The reference architecture of MIR.

SVF TVF SPR pretrained

TransMorph

VFA

Vector field attention: correspondences read directly from feature similarity. Light and contrast-robust.

SVF SPR pretrained

VFA: Vector Field Attention

SITReg

Multi-resolution, symmetric, inverse-consistent and topology-preserving by construction.

inverse consistent pretrained

SITReg

VoxelMorph

The U-Net baseline of learning-based registration.

displacement

VoxelMorph

LessNet

A very light convolutional network (0.1 M parameters).

SVF SPR

LessNet

Fourier-Net and Fourier-Net+

Band-limited deformations predicted in the Fourier domain; fast and compact.

SVF SPR

Fourier-Net and Fourier-Net+

Extensions

HyperTransMorph and HyperVFA

Hypernetwork-conditioned regularization strength, chosen at inference time.

HyperTransMorph

HyperMorph

The hypernetwork-conditioned VoxelMorph.

HyperMorph

LDDMM

Time-varying velocity fields with geodesic shooting, for TransMorph or any other backbone.

LDDMM

Optimization-based methods

ConvexAdam and FireANTs

Discrete convex optimization with Adam refinement, and GPU greedy / SyN registration.

SPR no training

ConvexAdam and FireANTs

deedsBCV

MRF-based discrete registration with self-similarity descriptors (Linux).

no training

deedsBCV

Atlases and building blocks

Atlas construction

Template building with any registration network, and MultiMorph group-wise atlases.

Atlas construction: MultiMorph and TemplateCreation

Backbones and heads

Swin encoders, decoders, segmentation, synthesis and self-supervised heads.

Backbones and heads

Comparison

Parameter counts are for the configurations used in the LUMIR25 tutorials (160 x 224 x 192 brain MRI; TransMorph runs at half resolution).

Model	Architecture	Parameters	Transformation models	Weights
TransMorph-TVF	Swin Transformer + conv decoder	17.1 M	displacement, SVF, TVF	LUMIR, IXI, autoPET
VFA	conv encoder + vector field attention	2.0 M	displacement, SVF	LUMIR, IXI
SITReg	multi-resolution symmetric	12.9 M	inverse consistent	IXI
VoxelMorph	U-Net	0.3 M	displacement	–
LessNet	light conv U-Net	0.1 M	displacement, SVF	–
Fourier-Net+	band-limited conv	0.9 M	displacement, SVF	–
ConvexAdam, deedsBCV, FireANTs	optimization	–	discrete, SVF, greedy / SyN	not needed

Every learned model above also accepts the LDDMM transformation model through LDDMMWrapper.