VBoussot commited on
Commit
6f387be
·
verified ·
1 Parent(s): 1396c9d

The benchmark is fidopt

Browse files
ANTs_SyN/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "ANTs SyN",
3
- "short_description": "SyN as the Learn2Reg 2024 organisers ran it for LUMIR (antsRegistrationSyN[so]): cross-correlation of radius 2, from the identity, 100x70x50x20 iterations, single precision, on the ANTs 2.6.5 release binaries fetched on first use. The ImagingBenchmarks baseline <code>ants</code>. CPU.",
4
- "description": "SyN as the Learn2Reg 2024 organisers ran it for LUMIR (antsRegistrationSyN[so]): cross-correlation of radius 2, from the identity, 100x70x50x20 iterations, single precision, on the ANTs 2.6.5 release binaries fetched on first use. The ImagingBenchmarks baseline <code>ants</code>. CPU. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "ANTs SyN",
3
+ "short_description": "SyN as the Learn2Reg 2024 organisers ran it for LUMIR (antsRegistrationSyN[so]): cross-correlation of radius 2, from the identity, 100x70x50x20 iterations, single precision, on the ANTs 2.6.5 release binaries fetched on first use. The fidopt baseline <code>ants</code>. CPU.",
4
+ "description": "SyN as the Learn2Reg 2024 organisers ran it for LUMIR (antsRegistrationSyN[so]): cross-correlation of radius 2, from the identity, 100x70x50x20 iterations, single precision, on the ANTs 2.6.5 release binaries fetched on first use. The fidopt baseline <code>ants</code>. CPU. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
Anatomix_ConvexAdam/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "anatomix + ConvexAdam",
3
- "short_description": "The ICLR 2025 pipeline of neel-dey/anatomix at a pinned commit, through its own convex_adam() with the defaults of its script: 16 anatomix and 12 MIND-SSC channels compared by SSD, diffusion weight 0.75, convex grid 2 with search half-width 1, 80 Adam steps on a grid of 2; CT clipped to [-450, 450] HU. The ImagingBenchmarks baseline <code>anatomix_convexadam</code>.",
4
- "description": "The ICLR 2025 pipeline of neel-dey/anatomix at a pinned commit, through its own convex_adam() with the defaults of its script: 16 anatomix and 12 MIND-SSC channels compared by SSD, diffusion weight 0.75, convex grid 2 with search half-width 1, 80 Adam steps on a grid of 2; CT clipped to [-450, 450] HU. The ImagingBenchmarks baseline <code>anatomix_convexadam</code>. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. The weights come from the authors' Hugging Face repository at a pinned revision. A GPU is needed. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "anatomix + ConvexAdam",
3
+ "short_description": "The ICLR 2025 pipeline of neel-dey/anatomix at a pinned commit, through its own convex_adam() with the defaults of its script: 16 anatomix and 12 MIND-SSC channels compared by SSD, diffusion weight 0.75, convex grid 2 with search half-width 1, 80 Adam steps on a grid of 2; CT clipped to [-450, 450] HU. The fidopt baseline <code>anatomix_convexadam</code>.",
4
+ "description": "The ICLR 2025 pipeline of neel-dey/anatomix at a pinned commit, through its own convex_adam() with the defaults of its script: 16 anatomix and 12 MIND-SSC channels compared by SSD, diffusion weight 0.75, convex grid 2 with search half-width 1, 80 Adam steps on a grid of 2; CT clipped to [-450, 450] HU. The fidopt baseline <code>anatomix_convexadam</code>. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. The weights come from the authors' Hugging Face repository at a pinned revision. A GPU is needed. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
BrainMorph/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "BrainMorph",
3
- "short_description": "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. The ImagingBenchmarks baseline <code>brainmorph</code>.",
4
- "description": "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. The ImagingBenchmarks baseline <code>brainmorph</code>. Brain MR, aligned pairs. <a href='https://github.com/alanqrwang/brainmorph/tree/106b029d6048d3d9d2f73cf7996a0b0813cbc4ae'>The authors' repository</a> at the commit of the LUMIR call (later commits dropped the H model), with its dependencies at versions of that date (torch 2.3 on Python 3.11) in an environment uv builds on first use (5 GB), and the weights checked against their hash. A GPU is recommended, up to the Ada generation: torch 2.3 has no kernels for Blackwell GPUs, so there the app runs on the CPU (no <code>--gpu</code>). Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).<br><br><b>How to cite:</b><br><cite>A. Q. Wang et al., <i>BrainMorph: A foundational keypoint model for robust and flexible brain MRI registration</i>, MELBA 2025.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "BrainMorph",
3
+ "short_description": "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. The fidopt baseline <code>brainmorph</code>.",
4
+ "description": "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. The fidopt baseline <code>brainmorph</code>. Brain MR, aligned pairs. <a href='https://github.com/alanqrwang/brainmorph/tree/106b029d6048d3d9d2f73cf7996a0b0813cbc4ae'>The authors' repository</a> at the commit of the LUMIR call (later commits dropped the H model), with its dependencies at versions of that date (torch 2.3 on Python 3.11) in an environment uv builds on first use (5 GB), and the weights checked against their hash. A GPU is recommended, up to the Ada generation: torch 2.3 has no kernels for Blackwell GPUs, so there the app runs on the CPU (no <code>--gpu</code>). Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).<br><br><b>How to cite:</b><br><cite>A. Q. Wang et al., <i>BrainMorph: A foundational keypoint model for robust and flexible brain MRI registration</i>, MELBA 2025.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
ConvexAdam_MIND/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "ConvexAdam (MIND), the authors' defaults",
3
- "short_description": "itk-impact's ConvexAdam (C++) at the settings of the authors' convex_adam_MIND: MIND-SSC features, a coupled-convex coarse search, then Adam instance optimisation, no linear stage. The ImagingBenchmarks baseline <code>ConvexAdam_MIND</code>.",
4
- "description": "itk-impact's ConvexAdam (C++) at the settings of the authors' convex_adam_MIND: MIND-SSC features, a coupled-convex coarse search, then Adam instance optimisation, no linear stage. The ImagingBenchmarks baseline <code>ConvexAdam_MIND</code>. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects). Settings as in the authors' repository (multimodallearning/convexAdam): control grid of 6 voxels, a discrete search of half-width 4, an SSD cost (L2 over the MIND channels), 80 Adam steps on a grid of 2 with a diffusion weight of 1.25 on the raw cost (no loss normalisation), three smoothing passes, and a learning rate of 2 in the image's voxels, which is their 1 in the half-resolution grid's. A native port of the authors' method, not their code: on the benchmark's check case it comes within 0.005 Dice of the authors' package, the rest being border conventions (zero padding of the search, the sampling of the inverse-consistency step, the upsampling of the coarse grid). Masks are ignored by this baseline. Output: the transform on the fixed grid.<br><br><b>How to cite:</b><br><cite>H. Siebert et al., <i>ConvexAdam: Self-configuring dual-optimisation-based 3D multitask medical image registration</i>, IEEE TMI 2024; V. Boussot et al., <i>IMPACT</i>, arXiv:2503.24121.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "ConvexAdam (MIND), the authors' defaults",
3
+ "short_description": "itk-impact's ConvexAdam (C++) at the settings of the authors' convex_adam_MIND: MIND-SSC features, a coupled-convex coarse search, then Adam instance optimisation, no linear stage. The fidopt baseline <code>ConvexAdam_MIND</code>.",
4
+ "description": "itk-impact's ConvexAdam (C++) at the settings of the authors' convex_adam_MIND: MIND-SSC features, a coupled-convex coarse search, then Adam instance optimisation, no linear stage. The fidopt baseline <code>ConvexAdam_MIND</code>. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects). Settings as in the authors' repository (multimodallearning/convexAdam): control grid of 6 voxels, a discrete search of half-width 4, an SSD cost (L2 over the MIND channels), 80 Adam steps on a grid of 2 with a diffusion weight of 1.25 on the raw cost (no loss normalisation), three smoothing passes, and a learning rate of 2 in the image's voxels, which is their 1 in the half-resolution grid's. A native port of the authors' method, not their code: on the benchmark's check case it comes within 0.005 Dice of the authors' package, the rest being border conventions (zero padding of the search, the sampling of the inverse-consistency step, the upsampling of the coarse grid). Masks are ignored by this baseline. Output: the transform on the fixed grid.<br><br><b>How to cite:</b><br><cite>H. Siebert et al., <i>ConvexAdam: Self-configuring dual-optimisation-based 3D multitask medical image registration</i>, IEEE TMI 2024; V. Boussot et al., <i>IMPACT</i>, arXiv:2503.24121.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
CorrField/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "corrField",
3
- "short_description": "corrField as the Lung250M-4B authors ship it, at a pinned commit, with its defaults: Foerstner keypoints inside the fixed image's lung mask, MIND-SSC, discrete matching, thin-plate splines. The ImagingBenchmarks baseline <code>corrfield</code>, lungs only: the fixed mask is required.",
4
- "description": "corrField as the Lung250M-4B authors ship it, at a pinned commit, with its defaults: Foerstner keypoints inside the fixed image's lung mask, MIND-SSC, discrete matching, thin-plate splines. The ImagingBenchmarks baseline <code>corrfield</code>, lungs only: the fixed mask is required. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. The pair is brought to a grid of at least 1 mm and at most 50 million voxels, and the field back onto the fixed grid. A GPU is needed. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "corrField",
3
+ "short_description": "corrField as the Lung250M-4B authors ship it, at a pinned commit, with its defaults: Foerstner keypoints inside the fixed image's lung mask, MIND-SSC, discrete matching, thin-plate splines. The fidopt baseline <code>corrfield</code>, lungs only: the fixed mask is required.",
4
+ "description": "corrField as the Lung250M-4B authors ship it, at a pinned commit, with its defaults: Foerstner keypoints inside the fixed image's lung mask, MIND-SSC, discrete matching, thin-plate splines. The fidopt baseline <code>corrfield</code>, lungs only: the fixed mask is required. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. The pair is brought to a grid of at least 1 mm and at most 50 million voxels, and the field back onto the fixed grid. A GPU is needed. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
Deeds/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "deeds",
3
- "short_description": "deedsBCV with its defaults: 5 levels, grid 8x7x6x5x4, search radius 8x7x6x5x4, quantisation 5x4x3x2x1, regularisation 1.6, symmetric; built from the authors' repository at a pinned commit on first use (make and a C++ compiler needed). The ImagingBenchmarks baseline <code>deeds</code>. CPU.",
4
- "description": "deedsBCV with its defaults: 5 levels, grid 8x7x6x5x4, search radius 8x7x6x5x4, quantisation 5x4x3x2x1, regularisation 1.6, symmetric; built from the authors' repository at a pinned commit on first use (make and a C++ compiler needed). The ImagingBenchmarks baseline <code>deeds</code>. CPU. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "deeds",
3
+ "short_description": "deedsBCV with its defaults: 5 levels, grid 8x7x6x5x4, search radius 8x7x6x5x4, quantisation 5x4x3x2x1, regularisation 1.6, symmetric; built from the authors' repository at a pinned commit on first use (make and a C++ compiler needed). The fidopt baseline <code>deeds</code>. CPU.",
4
+ "description": "deedsBCV with its defaults: 5 levels, grid 8x7x6x5x4, search radius 8x7x6x5x4, quantisation 5x4x3x2x1, regularisation 1.6, symmetric; built from the authors' repository at a pinned commit on first use (make and a C++ compiler needed). The fidopt baseline <code>deeds</code>. CPU. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
Demons/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "Demons",
3
- "short_description": "Demons as SimpleITK's registration notebook runs it: fast symmetric forces, 20 iterations, the field smoothed with a standard deviation of 2, two coarse levels then full resolution, histogram matching first for MR. One modality only. The ImagingBenchmarks baseline <code>demons</code>.",
4
- "description": "Demons as SimpleITK's registration notebook runs it: fast symmetric forces, 20 iterations, the field smoothed with a standard deviation of 2, two coarse levels then full resolution, histogram matching first for MR. One modality only. The ImagingBenchmarks baseline <code>demons</code>. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. SimpleITK (CPU). Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "Demons",
3
+ "short_description": "Demons as SimpleITK's registration notebook runs it: fast symmetric forces, 20 iterations, the field smoothed with a standard deviation of 2, two coarse levels then full resolution, histogram matching first for MR. One modality only. The fidopt baseline <code>demons</code>.",
4
+ "description": "Demons as SimpleITK's registration notebook runs it: fast symmetric forces, 20 iterations, the field smoothed with a standard deviation of 2, two coarse levels then full resolution, histogram matching first for MR. One modality only. The fidopt baseline <code>demons</code>. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. SimpleITK (CPU). Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
Greedy/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "greedy",
3
- "short_description": "greedy as its Quick Start runs it: normalised cross-correlation over a 2x2x2 window, 100x50x10 iterations, the deformable stage alone; the picsl_greedy package at a pinned version. One modality only. The ImagingBenchmarks baseline <code>greedy</code>. CPU.",
4
- "description": "greedy as its Quick Start runs it: normalised cross-correlation over a 2x2x2 window, 100x50x10 iterations, the deformable stage alone; the picsl_greedy package at a pinned version. One modality only. The ImagingBenchmarks baseline <code>greedy</code>. CPU. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "greedy",
3
+ "short_description": "greedy as its Quick Start runs it: normalised cross-correlation over a 2x2x2 window, 100x50x10 iterations, the deformable stage alone; the picsl_greedy package at a pinned version. One modality only. The fidopt baseline <code>greedy</code>. CPU.",
4
+ "description": "greedy as its Quick Start runs it: normalised cross-correlation over a 2x2x2 window, 100x50x10 iterations, the deformable stage alone; the picsl_greedy package at a pinned version. One modality only. The fidopt baseline <code>greedy</code>. CPU. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
MultiGradICON/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "multiGradICON",
3
- "short_description": "The multiGradICON network's prediction, then 50 steps of instance optimisation on LNCC: <code>unigradicon-register --model multigradicon</code> with its defaults, the ImagingBenchmarks baseline <code>multigradicon</code>.",
4
- "description": "The multiGradICON network's prediction, then 50 steps of instance optimisation on LNCC: <code>unigradicon-register --model multigradicon</code> with its defaults, the ImagingBenchmarks baseline <code>multigradicon</code>. With <code>--set masks=true</code> each image is multiplied by its lung mask, the authors' Learn2Reg lung protocol (the ImagingBenchmarks line <code>multigradicon_lungmask</code>). The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. The command knows two modalities, <code>ct</code> (clamped to [-1000, 1000] HU) and <code>mri</code> (minimum to the 99th percentile), read off the intensities. The weights come from <a href='https://github.com/uncbiag/uniGradICON'>the authors' release</a>, checked against their hash; the package is pinned at 1.0.5. A GPU is needed. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).<br><br><b>How to cite:</b><br><cite>L. Tian et al., <i>uniGradICON: A foundation model for medical image registration</i>, MICCAI 2024; B. Demir et al., <i>multiGradICON</i>, WBIR 2024.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "multiGradICON",
3
+ "short_description": "The multiGradICON network's prediction, then 50 steps of instance optimisation on LNCC: <code>unigradicon-register --model multigradicon</code> with its defaults, the fidopt baseline <code>multigradicon</code>.",
4
+ "description": "The multiGradICON network's prediction, then 50 steps of instance optimisation on LNCC: <code>unigradicon-register --model multigradicon</code> with its defaults, the fidopt baseline <code>multigradicon</code>. With <code>--set masks=true</code> each image is multiplied by its lung mask, the authors' Learn2Reg lung protocol (the fidopt line <code>multigradicon_lungmask</code>). The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. The command knows two modalities, <code>ct</code> (clamped to [-1000, 1000] HU) and <code>mri</code> (minimum to the 99th percentile), read off the intensities. The weights come from <a href='https://github.com/uncbiag/uniGradICON'>the authors' release</a>, checked against their hash; the package is pinned at 1.0.5. A GPU is needed. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).<br><br><b>How to cite:</b><br><cite>L. Tian et al., <i>uniGradICON: A foundation model for medical image registration</i>, MICCAI 2024; B. Demir et al., <i>multiGradICON</i>, WBIR 2024.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
NiftyReg/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "NiftyReg",
3
- "short_description": "reg_f3d with its defaults: cubic B-spline with control points every 5 voxels, NMI with 64 bins, bending energy 0.001, first-order penalty 0.01, no Jacobian penalty, 3 levels; NiftyReg v2.0.0 built from the authors' repository on first use (cmake and a C++ compiler needed). The ImagingBenchmarks baseline <code>niftyreg</code>. CPU.",
4
- "description": "reg_f3d with its defaults: cubic B-spline with control points every 5 voxels, NMI with 64 bins, bending energy 0.001, first-order penalty 0.01, no Jacobian penalty, 3 levels; NiftyReg v2.0.0 built from the authors' repository on first use (cmake and a C++ compiler needed). The ImagingBenchmarks baseline <code>niftyreg</code>. CPU. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "NiftyReg",
3
+ "short_description": "reg_f3d with its defaults: cubic B-spline with control points every 5 voxels, NMI with 64 bins, bending energy 0.001, first-order penalty 0.01, no Jacobian penalty, 3 levels; NiftyReg v2.0.0 built from the authors' repository on first use (cmake and a C++ compiler needed). The fidopt baseline <code>niftyreg</code>. CPU.",
4
+ "description": "reg_f3d with its defaults: cubic B-spline with control points every 5 voxels, NMI with 64 bins, bending energy 0.001, first-order penalty 0.01, no Jacobian penalty, 3 levels; NiftyReg v2.0.0 built from the authors' repository on first use (cmake and a C++ compiler needed). The fidopt baseline <code>niftyreg</code>. CPU. The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
README.md CHANGED
@@ -58,7 +58,7 @@ impact-reg-konfai register ConvexAdam_Composite -f fixed.nii.gz -m moving.nii.gz
58
 
59
  ## 📏 Baselines
60
 
61
- The registration baselines of [ImagingBenchmarks](https://github.com/fideus-labs/ImagingBenchmarks) live here too,
62
  next to the presets: each one runs its authors' code at a pinned version, fetched on first use, with the benchmark's
63
  settings in its `Prediction.yml`, and was checked against the benchmark's own line before it was published. Each app folder carries its code
64
  (`model.py`, with `baselines.py` for what the apps share, copied by `build.py`); `ConvexAdam_MIND` is a preset of
 
58
 
59
  ## 📏 Baselines
60
 
61
+ The registration baselines of [fidopt](https://github.com/fideus-labs/fidopt) live here too,
62
  next to the presets: each one runs its authors' code at a pinned version, fetched on first use, with the benchmark's
63
  settings in its `Prediction.yml`, and was checked against the benchmark's own line before it was published. Each app folder carries its code
64
  (`model.py`, with `baselines.py` for what the apps share, copied by `build.py`); `ConvexAdam_MIND` is a preset of
SynthMorph/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "SynthMorph",
3
- "short_description": "FreeSurfer's <code>mri_synthmorph register</code> with its defaults (regularisation 0.5, 7 integration steps, extent 256), the deformable model alone: a network trained on synthetic brain images of any contrast. The ImagingBenchmarks baseline <code>synthmorph</code>.",
4
- "description": "FreeSurfer's <code>mri_synthmorph register</code> with its defaults (regularisation 0.5, 7 integration steps, extent 256), the deformable model alone: a network trained on synthetic brain images of any contrast. The ImagingBenchmarks baseline <code>synthmorph</code>. Brain MR, aligned pairs: the tool's help says the deformable model assumes prior affine alignment. What the official image <code>freesurfer/synthmorph</code> holds, without a container: the script and its package from <a href='https://github.com/freesurfer/freesurfer/tree/21b31bf1b58ac195567c7777b21228130e4326e1/mri_synthmorph'>FreeSurfer's repository</a> at a pinned commit, the weights (3.5 GB) checked against their hash, and TensorFlow 2.17 on Python 3.12 in an environment uv builds on first use (7 GB). A GPU is recommended, up to the Ada generation: TensorFlow 2.17 has no kernels for Blackwell GPUs, so there the app runs on the CPU (no <code>--gpu</code>). Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).<br><br><b>How to cite:</b><br><cite>M. Hoffmann et al., <i>Anatomy-aware and acquisition-agnostic joint registration with SynthMorph</i>, Imaging Neuroscience 2024.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "SynthMorph",
3
+ "short_description": "FreeSurfer's <code>mri_synthmorph register</code> with its defaults (regularisation 0.5, 7 integration steps, extent 256), the deformable model alone: a network trained on synthetic brain images of any contrast. The fidopt baseline <code>synthmorph</code>.",
4
+ "description": "FreeSurfer's <code>mri_synthmorph register</code> with its defaults (regularisation 0.5, 7 integration steps, extent 256), the deformable model alone: a network trained on synthetic brain images of any contrast. The fidopt baseline <code>synthmorph</code>. Brain MR, aligned pairs: the tool's help says the deformable model assumes prior affine alignment. What the official image <code>freesurfer/synthmorph</code> holds, without a container: the script and its package from <a href='https://github.com/freesurfer/freesurfer/tree/21b31bf1b58ac195567c7777b21228130e4326e1/mri_synthmorph'>FreeSurfer's repository</a> at a pinned commit, the weights (3.5 GB) checked against their hash, and TensorFlow 2.17 on Python 3.12 in an environment uv builds on first use (7 GB). A GPU is recommended, up to the Ada generation: TensorFlow 2.17 has no kernels for Blackwell GPUs, so there the app runs on the CPU (no <code>--gpu</code>). Masks are ignored by this baseline. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).<br><br><b>How to cite:</b><br><cite>M. Hoffmann et al., <i>Anatomy-aware and acquisition-agnostic joint registration with SynthMorph</i>, Imaging Neuroscience 2024.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
UniGradICON/app.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "display_name": "uniGradICON",
3
- "short_description": "The uniGradICON network's prediction, then 50 steps of instance optimisation on LNCC: <code>unigradicon-register</code> with its defaults, the ImagingBenchmarks baseline <code>unigradicon</code>.",
4
- "description": "The uniGradICON network's prediction, then 50 steps of instance optimisation on LNCC: <code>unigradicon-register</code> with its defaults, the ImagingBenchmarks baseline <code>unigradicon</code>. With <code>--set masks=true</code> each image is multiplied by its lung mask, the authors' Learn2Reg lung protocol (the ImagingBenchmarks line <code>unigradicon_lungmask</code>). The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. The command knows two modalities, <code>ct</code> (clamped to [-1000, 1000] HU) and <code>mri</code> (minimum to the 99th percentile), read off the intensities. The weights come from <a href='https://github.com/uncbiag/uniGradICON'>the authors' release</a>, checked against their hash; the package is pinned at 1.0.5. A GPU is needed. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).<br><br><b>How to cite:</b><br><cite>L. Tian et al., <i>uniGradICON: A foundation model for medical image registration</i>, MICCAI 2024; B. Demir et al., <i>multiGradICON</i>, WBIR 2024.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,
 
1
  {
2
  "display_name": "uniGradICON",
3
+ "short_description": "The uniGradICON network's prediction, then 50 steps of instance optimisation on LNCC: <code>unigradicon-register</code> with its defaults, the fidopt baseline <code>unigradicon</code>.",
4
+ "description": "The uniGradICON network's prediction, then 50 steps of instance optimisation on LNCC: <code>unigradicon-register</code> with its defaults, the fidopt baseline <code>unigradicon</code>. With <code>--set masks=true</code> each image is multiplied by its lung mask, the authors' Learn2Reg lung protocol (the fidopt line <code>unigradicon_lungmask</code>). The pair must be aligned (rigidly for one patient, by an affine transform between two subjects): only the deformable stage runs. The command knows two modalities, <code>ct</code> (clamped to [-1000, 1000] HU) and <code>mri</code> (minimum to the 99th percentile), read off the intensities. The weights come from <a href='https://github.com/uncbiag/uniGradICON'>the authors' release</a>, checked against their hash; the package is pinned at 1.0.5. A GPU is needed. Output: the displacement field on the fixed grid, in mm along the world axes (ITK, LPS).<br><br><b>How to cite:</b><br><cite>L. Tian et al., <i>uniGradICON: A foundation model for medical image registration</i>, MICCAI 2024; B. Demir et al., <i>multiGradICON</i>, WBIR 2024.</cite>",
5
  "task": "registration",
6
  "tta": 0,
7
  "mc_dropout": 0,