Each one runs its authors' code at a fixed version, from the same starting point as the others. Get its app, or run it on your own machine.
41 methods · sorted by Name Datasets ranked on
A KonfAI segmentation app run through konfai-apps: the variant names an app of fideus-labs/fidopt-segmentation, or one of Valentin Boussot's apps where it lives; the app writes its own labels, matched with the dataset's structures by name.
convex_adam_MIND with the repository's defaults: MIND-SSC features, a coupled-convex coarse search, then Adam instance optimisation.
ITK's fast symmetric forces Demons as SimpleITK's Demons notebook runs it: 20 iterations, field smoothing 2, pyramid 4, 2, 1; on MR the moving histogram is first matched to the fixed one, as SimpleITK's Demons examples do.
MONAI's wholeBody_ct_segmentation bundle, a SegResNet trained on the TotalSegmentator dataset at 1.5 mm: TotalSegmentator v1's 104 structures of a CT, run as the bundle's inference.json runs it.
reg_f3d with its defaults (cubic B-spline, NMI), from the corpus's rigidly aligned pair.
CT-FM's whole-body model, a SegResNet pretrained by self-supervision on 148,000 CTs and fine-tuned on TotalSegmentator's training split: TotalSegmentator v2's 117 structures of a CT at its own resolution, run as its model card's notebook runs it.
greedy's deformable Quick Start command as first written (2019): NCC on a 2x2x2 window, 100x50x10 iterations, default smoothing; one intensity space, as its reference advises (NMI does not work very well for deformable registration).
2D feature matchers run by vismatch, each with its authors' code and weights, the homography from the fixed image to the moving one fitted by OpenCV's MAGSAC: the variant names the matcher.
deedsBCV with its defaults (MIND-SSC, discrete optimisation), from the corpus's rigidly aligned pair, on an isotropic grid as its README asks for thick slices.
MONAI's brats_mri_segmentation bundle, a SegResNet trained on BraTS 2018: the tumour core, whole tumour and enhancing tumour of a glioma from T1, T1c, T2 and FLAIR, as BraTS labels, run as the bundle's inference.json runs it.
ConvexAdam on anatomix and MIND-SSC features, the registration pipeline of the anatomix paper with its command-line defaults.
lungmask's U-Net(R231), trained on 231 routine CTs of varied disease, the right and left lungs slice by slice, run as its authors' command line runs it.
FastSurferVINN, three voxel-size independent 2D networks (axial, coronal, sagittal) whose views are aggregated: 95 structures of a T1-weighted brain, FreeSurfer's aseg and the 62 DKT cortical regions, run as its authors' run_fastsurfer.sh --seg_only runs it.
unigradicon-register --model multigradicon with its defaults: the network's prediction, then 50 steps of instance optimisation on LNCC.
SynthSeg 1.0, a U-Net trained only on synthetic scans drawn from brain label maps, so that one model reads any MR contrast and resolution: 32 brain structures, run as its authors' SynthSeg_predict.py --v1 runs it.
SyN as the Learn2Reg 2024 organisers ran it for LUMIR: cross-correlation, from the identity, 100x70x50x20 iterations, single precision.
DFCAN, the Fourier channel attention network of the BioSR paper, with its authors' single-image models (wide field to linear SIM, one model a structure), run as their predict.py runs it.
A classical baseline with no learnt weights: Otsu's threshold with scikit-image: the nuclei above it on a stained section's haematoxylin (colour deconvolution), or on the side of it that holds fewer pixels on a grey image; small holes and specks removed.
SynthSR 1.0, a U-Net trained on scans synthesised from brain label maps to turn a brain MR of any contrast, orientation and resolution into a 1 mm MPRAGE-like T1, run as its authors' predict_command_line.py runs it; brought back onto the input's grid and scaled onto the input's intensities.
unigradicon-register with its defaults: the network's prediction, then 50 steps of instance optimisation on LNCC.
ANTs' DenoiseImage, Manjón et al.'s adaptive non-local means for MR, at the command's defaults (Gaussian noise model, patch radius 1, search radius 2): no learnt weights.
corrfield.py with its defaults, on the authors' 1 mm grid or coarser: Foerstner keypoints inside the fixed image's lung mask, MIND-SSC, discrete matching, thin-plate splines.
DIPY's blockwise non-local means with a Rician noise model, the noise's level estimated from the image by estimate_sigma, both at their defaults: no learnt weights.
SimpleITK's recursive Gaussian smoothing at its default standard deviation, 1 mm along every axis: the simplest denoiser, no learnt weights.
InstanSeg's released nuclei models, brightfield (trained on TNBC, LyNSeC, NuInsSeg, IHC TMA and CoNSeP) or fluorescence (CPDMI 2023) as the image calls for, run as the package's README runs them, every nucleus one label.
Little W-Net, a small W-shaped network for retinal vessels, with its authors' DRIVE model, run by their predict_one_image.py: test-time flips, the threshold 0.42 they chose on DRIVE.
BrainMorph as the LUMIR organisers run it: 512 keypoints, the H model, thin-plate splines without regularisation; a keypoint network trained on 100,000 brain MR images.
Cellpose 3's generalist cyto3 model, trained on nine datasets among which the 2018 Data Science Bowl, run as the package's Cellpose class runs it at its defaults (grey, diameter 30), every mask one nucleus.
elastix's own default B-spline parameter map (mutual information with a bending-energy penalty, four resolutions, 256 iterations, a 10 mm grid), run by IMPACT-Reg's elastix engine with the IMPACT metric unused: a preset of that engine, not elastix's own executable.
FCBFormer, a PVTv2 transformer branch fused with a fully convolutional branch, with its authors' Kvasir-SEG model, run as their predict.py runs it: 352 x 352, the logits kept above 0.
The FireANTs library used alone as its first tutorial does: greedy deformable registration on local cross-correlation, from the aligned pair (the tutorial's affine stage left out, as for every method). Kept for the comparison of costs with IMPACT-Reg's FireANTs engine.
A KonfAI synthesis app run through konfai-apps: the variant names an app of fideus-labs/fidopt-synthesis, or one of Valentin Boussot's apps where it lives; the app writes the image of the target modality on the image's grid.
SAM 3 fine-tuned on 33 medical datasets, its authors' 2D model prompted by text alone, no box or point: the structure the authors' prompt names, the mask of highest score above 0.1, as their evaluation script runs it.
multiGradICON with its authors' lung protocol: each image multiplied by its lung mask. The field means nothing outside the lungs.
nnsyn, CSIRO AEHRC's nnU-Net ResEnc-L that regresses the CT (team KoalAI, first in SynthRAD2025's MR-to-CT task), with its authors' released SynthRAD2025 models, one per source modality and region, run as their challenge docker runs them on the one fold they published.
StarDist's pretrained 2D_versatile_he, nuclei as star-convex polygons, trained on MoNuSeg 2018 and TNBC, run as the README runs a pretrained model, every nucleus one label.
SynthMorph, networks trained only on synthetic images drawn from brain label maps, so that one model registers brain MR of any contrast and resolution without preprocessing: joint affine and deformable registration, run as its authors' mri_synthmorph register runs it.
TorchXRayVision's chest radiograph anatomy model, a PSPNet whose training images were not released: 14 structures, lungs and heart among them, from a frontal radiograph, run as the authors' segmentation notebook runs it.
uniGradICON with its authors' lung protocol: each image multiplied by its lung mask. The field means nothing outside the lungs.
VM-UNet, a U-shaped network of Vision Mamba blocks, with its authors' ISIC 2018 checkpoint, run as their test loop runs it: normalised with their ISIC 2018 statistics, resized to 256 x 256, the sigmoid kept above 0.5.
The warpfield library with its default recipe's two deformable levels: FFT block matching of 2D projections on the GPU, from the aligned pair (the recipe's translation and affine levels left out, as for every method). The moving image is brought onto the fixed grid; each image is shifted to start at zero, the recipe's background clipping being made for microscopy.
A method of yours can stand here too: how to submit.
fidopt