VBoussot
/

ImpactReg

Model card Files Files and versions

xet

Community

VBoussot commited on 20 days ago

Commit

566e8a2

verified ·

1 Parent(s): aeea6b4

Update PresetDatabase.json

Browse files

Files changed (1) hide show

PresetDatabase.json +18 -18

PresetDatabase.json CHANGED Viewed

@@ -32,9 +32,9 @@
         "ParameterMap_MRI_HN.txt"
       ],
       "models": [
-        "VBoussot/impact-torchscript-models:MIND/R1D2.pt"
       ],
-      "preprocess_function": "Preprocess:standardize_MRI",
       "iterations": 1050,
       "short_description": "Optimized preset for MR/CT registration on head & neck",
       "description": "A four-level recursive B-spline deformable registration optimized for MRI head-and-neck images, driven by the IMPACT metric and combining semantic features extracted from the pretrained MIND model at progressively finer voxel scales (6 mm, 3 mm, 2 mm, 2 mm). The optimization follows a multi-resolution scheme with up to 300, 300, 250, and 200 ASGD iterations, using stochastic sampling and a composite metric (IMPACT + mutual information + bending energy) to achieve robust semantic alignment in complex head-and-neck anatomy."
@@ -46,10 +46,10 @@
         "ParameterMap_MRI_MRSeg.txt"
       ],
       "models": [
-        "VBoussot/impact-torchscript-models:MRSeg/MRSeg_8_Layers.pt",
-        "VBoussot/impact-torchscript-models:MIND/R1D2.pt"
       ],
-      "preprocess_function": "Preprocess:standardize_MRI",
       "iterations": 1100,
       "short_description": "Generic MR/CT deformable registration using MIND + MRSegmentator features",
       "description": "A four-level recursive B-spline deformable registration optimized for generic MR/CT alignment, driven by the IMPACT metric and combining semantic features from two pretrained models: MIND (L1 distance on a subset of 32 features) and MRSegmentator (Dice overlap on segmentation outputs with 64 features). Features are extracted at progressively finer voxel scales (6 mm, 3 mm, 2 mm, 2 mm) with level-dependent weighting between MIND and MRSegmentator (0.2/0.8, 0.3/0.7, 0.6/0.4, 0.7/0.3). The optimization follows a multi-resolution ASGD scheme with up to 400, 300, 200, and 200 iterations using 2000 random spatial samples, and a composite objective (IMPACT + mutual information + bending energy penalty) to ensure robust cross-modality semantic alignment and smooth deformations."
@@ -61,10 +61,10 @@
         "ParameterMap_MRI_TS.txt"
       ],
       "models": [
-        "VBoussot/impact-torchscript-models:TS/M852_8_Layers.pt",
-        "VBoussot/impact-torchscript-models:MIND/R1D2.pt"
       ],
-      "preprocess_function": "Preprocess:standardize_MRI",
       "iterations": 1100,
       "short_description": "Generic MR/CT deformable registration using MIND + Totalsegmentator features",
       "description": "A four-level recursive B-spline deformable registration optimized for generic MR/CT alignment, driven by the IMPACT metric and combining semantic features from two pretrained models: MIND (L1 distance on a subset of 32 features) and Totalsegmentator (Dice overlap on segmentation outputs with 64 features). Features are extracted at progressively finer voxel scales (6 mm, 3 mm, 2 mm, 2 mm) with level-dependent weighting between MIND and MRSegmentator (0.2/0.8, 0.3/0.7, 0.6/0.4, 0.7/0.3). The optimization follows a multi-resolution ASGD scheme with up to 400, 300, 200, and 200 iterations using 2000 random spatial samples, and a composite objective (IMPACT + mutual information + bending energy penalty) to ensure robust cross-modality semantic alignment and smooth deformations."
@@ -76,10 +76,10 @@
         "ParameterMap_CBCT_generic_MRSeg.txt"
       ],
       "models": [
-        "VBoussot/impact-torchscript-models:MRSeg/MRSeg_8_Layers.pt",
-        "VBoussot/impact-torchscript-models:MIND/R1D2.pt"
       ],
-      "preprocess_function": "Preprocess:standardize_MRI",
       "iterations": 1050,
       "short_description": "Generic CBCT/CT deformable registration using MRSegmentator features",
       "description": "A four-level recursive B-spline deformable registration optimized for generic CBCT/CT alignment, driven by the IMPACT metric using semantic features extracted from the pretrained MRSegmentator model. The scheme follows a multi-resolution strategy with up to 300, 300, 250, and 200 ASGD iterations and 2000 stochastic spatial samples per level. Features are extracted at progressively finer voxel scales (3 mm, 3 mm, 2×2×3 mm, 2×2×3 mm), with a level-dependent combination of Dice-based segmentation overlap and L1 feature distances on selected internal layers of MRSegmentator. Early levels rely on pure Dice supervision, while finer stages progressively integrate feature-level alignment with increasing L1 contribution (0.3/0.7, 0.5/0.5) and a final purely feature-based stage. The optimization minimizes a composite objective (IMPACT + mutual information + bending energy penalty), enabling robust cross-modality alignment between CBCT and CT while enforcing smooth, physically plausible deformations."
@@ -91,10 +91,10 @@
         "ParameterMap_CBCT_generic_TS.txt"
       ],
       "models": [
-        "VBoussot/impact-torchscript-models:TS/M852_8_Layers.pt",
-        "VBoussot/impact-torchscript-models:TS/M850_8_Layers.pt"
       ],
-      "preprocess_function": "Preprocess:standardize_MRI",
       "iterations": 1050,
       "short_description": "Generic CBCT/CT deformable registration using TotalSegmentator features",
       "description": "A four-level recursive B-spline deformable registration optimized for generic CBCT/CT alignment, driven by the IMPACT metric using semantic features extracted from pretrained TotalSegmentator TorchScript models. The optimization follows a multi-resolution ASGD scheme with up to 300, 300, 250, and 200 iterations using 2000 random spatial samples per level. Features are extracted at progressively finer voxel scales (3 mm, 3 mm, 2×2×3 mm, 2×2×3 mm), starting with Dice-based overlap on segmentation outputs and progressively integrating feature-level alignment via L1 distances on selected internal layers (0.3/0.7 then 0.5/0.5 L1/Dice), ending with a final purely feature-based stage. A composite objective (IMPACT + mutual information + bending energy penalty) ensures robust cross-modality alignment while enforcing smooth, physically plausible deformations."
@@ -106,11 +106,11 @@
         "ParameterMap_CBCT_HN.txt"
       ],
       "models": [
-        "VBoussot/impact-torchscript-models:TS/M732_2_Layers.pt",
-        "VBoussot/impact-torchscript-models:TS/M731_2_Layers.pt",
-        "VBoussot/impact-torchscript-models:TS/M730_2_Layers.pt"
       ],
-      "preprocess_function": "Preprocess:standardize_MRI",
       "iterations": 1150,
       "short_description": "Optimized preset for CBCT/CT registration on head & neck",
       "description": "A five-level recursive B-spline deformable registration optimized for CBCT/CT head-and-neck alignment, driven by the IMPACT metric using semantic features extracted from pretrained TotalSegmentator TorchScript models. The optimization follows a multi-resolution ASGD scheme with up to 300, 300, 200, 200, and 150 iterations and 2000 stochastic spatial samples per level. Features are extracted at progressively finer voxel scales (6 mm, 3 mm, 3 mm, 2×2×3 mm, 2×2×3 mm) using L1 distances on selected internal layers of the network. A composite objective (IMPACT + mutual information + bending energy penalty, with increased MI weight) ensures robust cross-modality alignment in complex head-and-neck anatomy while enforcing smooth, physically plausible deformations."

         "ParameterMap_MRI_HN.txt"
       ],
       "models": [
+        "VBoussot/impact-torchscript-models:MIND/R1D2_3D.pt"
       ],
+      "preprocess_function": "",
       "iterations": 1050,
       "short_description": "Optimized preset for MR/CT registration on head & neck",
       "description": "A four-level recursive B-spline deformable registration optimized for MRI head-and-neck images, driven by the IMPACT metric and combining semantic features extracted from the pretrained MIND model at progressively finer voxel scales (6 mm, 3 mm, 2 mm, 2 mm). The optimization follows a multi-resolution scheme with up to 300, 300, 250, and 200 ASGD iterations, using stochastic sampling and a composite metric (IMPACT + mutual information + bending energy) to achieve robust semantic alignment in complex head-and-neck anatomy."
         "ParameterMap_MRI_MRSeg.txt"
       ],
       "models": [
+        "VBoussot/impact-torchscript-models:MRSeg/MRSeg.pt",
+        "VBoussot/impact-torchscript-models:MIND/R1D2_3D.pt"
       ],
+      "preprocess_function": "",
       "iterations": 1100,
       "short_description": "Generic MR/CT deformable registration using MIND + MRSegmentator features",
       "description": "A four-level recursive B-spline deformable registration optimized for generic MR/CT alignment, driven by the IMPACT metric and combining semantic features from two pretrained models: MIND (L1 distance on a subset of 32 features) and MRSegmentator (Dice overlap on segmentation outputs with 64 features). Features are extracted at progressively finer voxel scales (6 mm, 3 mm, 2 mm, 2 mm) with level-dependent weighting between MIND and MRSegmentator (0.2/0.8, 0.3/0.7, 0.6/0.4, 0.7/0.3). The optimization follows a multi-resolution ASGD scheme with up to 400, 300, 200, and 200 iterations using 2000 random spatial samples, and a composite objective (IMPACT + mutual information + bending energy penalty) to ensure robust cross-modality semantic alignment and smooth deformations."
         "ParameterMap_MRI_TS.txt"
       ],
       "models": [
+        "VBoussot/impact-torchscript-models:TS/M852.pt",
+        "VBoussot/impact-torchscript-models:MIND/R1D2_3D.pt"
       ],
+      "preprocess_function": "",
       "iterations": 1100,
       "short_description": "Generic MR/CT deformable registration using MIND + Totalsegmentator features",
       "description": "A four-level recursive B-spline deformable registration optimized for generic MR/CT alignment, driven by the IMPACT metric and combining semantic features from two pretrained models: MIND (L1 distance on a subset of 32 features) and Totalsegmentator (Dice overlap on segmentation outputs with 64 features). Features are extracted at progressively finer voxel scales (6 mm, 3 mm, 2 mm, 2 mm) with level-dependent weighting between MIND and MRSegmentator (0.2/0.8, 0.3/0.7, 0.6/0.4, 0.7/0.3). The optimization follows a multi-resolution ASGD scheme with up to 400, 300, 200, and 200 iterations using 2000 random spatial samples, and a composite objective (IMPACT + mutual information + bending energy penalty) to ensure robust cross-modality semantic alignment and smooth deformations."
         "ParameterMap_CBCT_generic_MRSeg.txt"
       ],
       "models": [
+        "VBoussot/impact-torchscript-models:MRSeg/MRSeg.pt",
+        "VBoussot/impact-torchscript-models:MIND/R1D2_3D.pt"
       ],
+      "preprocess_function": "",
       "iterations": 1050,
       "short_description": "Generic CBCT/CT deformable registration using MRSegmentator features",
       "description": "A four-level recursive B-spline deformable registration optimized for generic CBCT/CT alignment, driven by the IMPACT metric using semantic features extracted from the pretrained MRSegmentator model. The scheme follows a multi-resolution strategy with up to 300, 300, 250, and 200 ASGD iterations and 2000 stochastic spatial samples per level. Features are extracted at progressively finer voxel scales (3 mm, 3 mm, 2×2×3 mm, 2×2×3 mm), with a level-dependent combination of Dice-based segmentation overlap and L1 feature distances on selected internal layers of MRSegmentator. Early levels rely on pure Dice supervision, while finer stages progressively integrate feature-level alignment with increasing L1 contribution (0.3/0.7, 0.5/0.5) and a final purely feature-based stage. The optimization minimizes a composite objective (IMPACT + mutual information + bending energy penalty), enabling robust cross-modality alignment between CBCT and CT while enforcing smooth, physically plausible deformations."
         "ParameterMap_CBCT_generic_TS.txt"
       ],
       "models": [
+        "VBoussot/impact-torchscript-models:TS/M852.pt",
+        "VBoussot/impact-torchscript-models:TS/M850.pt"
       ],
+      "preprocess_function": "",
       "iterations": 1050,
       "short_description": "Generic CBCT/CT deformable registration using TotalSegmentator features",
       "description": "A four-level recursive B-spline deformable registration optimized for generic CBCT/CT alignment, driven by the IMPACT metric using semantic features extracted from pretrained TotalSegmentator TorchScript models. The optimization follows a multi-resolution ASGD scheme with up to 300, 300, 250, and 200 iterations using 2000 random spatial samples per level. Features are extracted at progressively finer voxel scales (3 mm, 3 mm, 2×2×3 mm, 2×2×3 mm), starting with Dice-based overlap on segmentation outputs and progressively integrating feature-level alignment via L1 distances on selected internal layers (0.3/0.7 then 0.5/0.5 L1/Dice), ending with a final purely feature-based stage. A composite objective (IMPACT + mutual information + bending energy penalty) ensures robust cross-modality alignment while enforcing smooth, physically plausible deformations."
         "ParameterMap_CBCT_HN.txt"
       ],
       "models": [
+        "VBoussot/impact-torchscript-models:TS/M732.pt",
+        "VBoussot/impact-torchscript-models:TS/M731.pt",
+        "VBoussot/impact-torchscript-models:TS/M730.pt"
       ],
+      "preprocess_function": "",
       "iterations": 1150,
       "short_description": "Optimized preset for CBCT/CT registration on head & neck",
       "description": "A five-level recursive B-spline deformable registration optimized for CBCT/CT head-and-neck alignment, driven by the IMPACT metric using semantic features extracted from pretrained TotalSegmentator TorchScript models. The optimization follows a multi-resolution ASGD scheme with up to 300, 300, 200, 200, and 150 iterations and 2000 stochastic spatial samples per level. Features are extracted at progressively finer voxel scales (6 mm, 3 mm, 3 mm, 2×2×3 mm, 2×2×3 mm) using L1 distances on selected internal layers of the network. A composite objective (IMPACT + mutual information + bending energy penalty, with increased MI weight) ensures robust cross-modality alignment in complex head-and-neck anatomy while enforcing smooth, physically plausible deformations."