TransMorph

Overview

TransMorph is the reference architecture of the toolkit: a Swin-Transformer encoder on the concatenated moving/fixed pair, a convolutional decoder with skip connections, and a registration head that predicts a dense field. It is accurate on brain MRI and whole-body PET/CT, trains with any similarity loss in MIR, and is the backbone of the spatially varying regularization, hypernetwork and LDDMM variants.

Variants

Class	Description	Output
TransMorph	Original model; SVF=True integrates the output as a stationary velocity field (SVF_steps scaling-and-squaring steps).	flow
TransMorphTVF	Time-varying field: time_steps per-step heads each predict a small displacement from the currently warped moving image; the displacements are composed (composition='composition') or added. The LUMIR24/25 brain models use this variant with time_steps=7.	flow
TransMorphTVFSPR	TransMorph-TVF with a spatially varying regularization head: returns the per-voxel weight map used by LocalGrad3d / logBeta; SVF=True also returns the inverse flow.	(flow, wts) or (pos_flow, neg_flow, wts)
TransMorphTVFAutoPETAtlas	PET/CT atlas-registration variant integrated at twice the network resolution.	(pos_flow, neg_flow)
TransMorphAffine	Affine registration head on the TransMorph encoder.	affine parameters
swin_type='dswin' / 'dswinv2'	Deformable cross-attention transformer (TransMorph-DCA) encoders.	as above
TransMorphLDDMM	TransMorph with the LDDMM transformation model (LDDMM).	flow

Usage

import torch.nn.functional as F
from MIR.models import TransMorphTVF, SpatialTransformer
import MIR.models.configs_TransMorph as configs_TransMorph

H, W, D = 160, 224, 192                       # LUMIR image size; the network runs at half resolution
config = configs_TransMorph.get_3DTransMorph3Lvl_config()
config.img_size = (H // 2, W // 2, D // 2)
config.window_size = (H // 64, W // 64, D // 64)
model = TransMorphTVF(config, time_steps=7).cuda()

flow_half = model((F.avg_pool3d(mov, 2), F.avg_pool3d(fix, 2)))
flow = F.interpolate(flow_half, scale_factor=2, mode='trilinear', align_corners=False) * 2
warped = SpatialTransformer((H, W, D)).cuda()(mov, flow)

Configurations live in MIR.models.configs_TransMorph (get_3DTransMorph_config, get_3DTransMorph3Lvl_config, get_3DTransMorphAtlas3Lvl_config, large / small / tiny presets, ablations of the skip connections and positional embeddings). config.img_size must be divisible by patch_size * 2 ** (levels - 1) and window_size by the feature-map size.

Pretrained weights

ModelWeights['TransMorphTVF-LUMIR24-MonoModal'] (brain, LUMIR24), ModelWeights['TransMorph-IXI-Atlas2Subject'], ModelWeights['TransMorphTVF-IXI-Atlas2Subject'], ModelWeights['MedIA-TM-SPR-Beta-autoPET'] and ['MedIA-TM-SPR-Gaussian-autoPET'] (TransMorph-TVF-SPR, PET/CT).

Key modules

• MIR.models.TransMorph, MIR.models.configs_TransMorph

• MIR.models.Swin_Transformer, MIR.models.Deformable_Swin_Transformer, MIR.models.Deformable_Swin_Transformer_v2

• MIR.models.HyperTransMorph (HyperTransMorph)

Related tutorials

• LUMIR25 TransMorph (training template)

• LUMIR24 TransMorph

• IXI benchmarking

• IXI TransMorph-SPR

• autoPET TransMorph-SPR

• Brain template construction

References

• Chen et al. TransMorph: Transformer for unsupervised medical image registration. Medical Image Analysis 2022. arXiv

• Chen et al. Unsupervised learning of diffeomorphic image registration via TransMorph. WBIR 2022. (TransMorph-TVF)

• Chen et al. Deformable cross-attention transformer for medical image registration. MLMI, MICCAI 2023. (TransMorph-DCA)

• Chen et al. Unsupervised learning of spatially varying regularization for diffeomorphic image registration. Medical Image Analysis 2025. (TransMorph-TVF-SPR)