Pretrained weights

Checkpoints are hosted on Google Drive and registered by name in MIR.ModelWeights. Each entry gives the Drive file id (wts) and the key under which the state dict is stored (wts_key).

Loading a checkpoint

import gdown, torch
from MIR import ModelWeights

entry = ModelWeights['TransMorphTVF-LUMIR24-MonoModal']
gdown.download(f"https://drive.google.com/uc?id={entry['wts']}", 'TransMorphTVF_LUMIR24.pth.tar')
model.load_state_dict(torch.load('TransMorphTVF_LUMIR24.pth.tar')[entry['wts_key']])

To initialize a modified architecture from a checkpoint, load only the tensors whose names and shapes match:

from MIR.utils import load_partial_weights
load_partial_weights(model, 'TransMorphTVF_LUMIR24.pth.tar', weights_key='state_dict')

Available checkpoints

ModelWeights key

Model

Trained on

TransMorphTVF-LUMIR24-MonoModal

TransMorph-TVF

LUMIR24 brain T1w MRI, inter-subject (Learn2Reg 2024)

VFA-LUMIR24-MonoModal

VFA

LUMIR24 brain T1w MRI (key model_state_dict)

VFA-LUMIR25-MultiModal

VFA

LUMIR25 brain MRI, multiple contrasts (Learn2Reg 2025)

TransMorph-IXI-Atlas2Subject

TransMorph

IXI brain MRI, atlas-to-subject

TransMorphTVF-IXI-Atlas2Subject

TransMorph-TVF

IXI brain MRI, atlas-to-subject

VFA-IXI-Atlas2Subject

VFA

IXI brain MRI, atlas-to-subject

VFA-SPR-IXI-Atlas2Subject

VFA-SPR

IXI brain MRI, atlas-to-subject

SITReg-IXI-Atlas2Subject

SITReg

IXI brain MRI, atlas-to-subject

SITReg-SPR-IXI-Atlas2Subject

SITReg-SPR

IXI brain MRI, atlas-to-subject

MedIA-TM-SPR-Beta-autoPET

TransMorph-TVF-SPR, Beta prior

autoPET whole-body PET/CT, atlas-to-subject (MedIA 2025)

MedIA-TM-SPR-Gaussian-autoPET

TransMorph-TVF-SPR, Gaussian prior

autoPET whole-body PET/CT, atlas-to-subject (MedIA 2025)

VFA-SegHead

VFA segmentation head

VFA features for segmentation

VFA-SynthHead

VFA synthesis head

VFA features for image synthesis

Datasets and atlases

MIR.DatasetJSONs holds the dataset split files of LUMIR24 and LUMIR25, MIR.ValEvalModules the official validation-evaluation modules, and MIR.pretrained_wts.FDG_atlas_autoPET an FDG PET/CT atlas built from autoPET (CT, SUV, organ probability and label atlases, 2.8 x 2.8 x 3.8 mm, sex-specific variants).