VBoussot commited on
Commit
1396c9d
·
verified ·
1 Parent(s): 8b2856d

ConvexAdam_MIND: the authors' cost, balance and step

Browse files

distance L2 (an SSD over the MIND channels, in the coarse search as in the Adam stage), normalize false (the raw similarity against the diffusion weight, as the authors balance them) and learning rate 2 in the image's voxels (their 1 in the half-resolution grid's). On the check case the port now comes within 0.005 Dice of the authors' package instead of 0.025; what remains is border conventions of the coarse stage.

ConvexAdam_MIND/Prediction.yml CHANGED
@@ -8,13 +8,13 @@ Predictor:
8
  layers_mask: '1'
9
  layers_weight: 1.0
10
  pca: 0
11
- distance: L1
12
  overlap: 2
13
  mixed_precision: false
14
  grid_spacing: 6
15
  displacement_half_width: 4
16
  iterations: 80
17
- learning_rate: 1.0
18
  regularization_weight: 1.25
19
  grid_shrink: 2
20
  control_grid_smoothing: 3
@@ -22,6 +22,7 @@ Predictor:
22
  linear: false
23
  linear_iterations: 200
24
  seed: 42
 
25
  outputs_criterions: None
26
  stages:
27
  - coarse
 
8
  layers_mask: '1'
9
  layers_weight: 1.0
10
  pca: 0
11
+ distance: L2
12
  overlap: 2
13
  mixed_precision: false
14
  grid_spacing: 6
15
  displacement_half_width: 4
16
  iterations: 80
17
+ learning_rate: 2.0
18
  regularization_weight: 1.25
19
  grid_shrink: 2
20
  control_grid_smoothing: 3
 
22
  linear: false
23
  linear_iterations: 200
24
  seed: 42
25
+ normalize: false
26
  outputs_criterions: None
27
  stages:
28
  - coarse
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, 80 Adam steps at learning rate 1 on a grid of 2 with a diffusion weight of 1.25 and three smoothing passes. A native port of the authors' method, not their code: on the benchmark's check case it comes within 0.03 Dice of the authors' package. 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 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,