Hugging Face's logo

TaouliLab
/
LiverSegMRI

Image Segmentation
nnunetv2
medical-imaging
MRI
liver
segmentation
nnU-Net
radiology
Model card Files
xet
 Community
TaouliLab commited on
Commit
cab169c
·
verified ·
1 Parent(s): 5972db1

LiverSegMRI weights (five folds, as trained), plans, and model card

Browse files
Files changed (8) hide show
  1. README.md +130 -0
  2. dataset.json +12 -0
  3. fold_0/checkpoint_final.pth +3 -0
  4. fold_1/checkpoint_final.pth +3 -0
  5. fold_2/checkpoint_final.pth +3 -0
  6. fold_3/checkpoint_final.pth +3 -0
  7. fold_4/checkpoint_final.pth +3 -0
  8. plans.json +521 -0
README.md ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ library_name: nnunetv2
4
+ pipeline_tag: image-segmentation
5
+ tags:
6
+ - medical-imaging
7
+ - MRI
8
+ - liver
9
+ - segmentation
10
+ - nnU-Net
11
+ - radiology
12
+ ---
13
+
14
+ # LiverSegMRI
15
+
16
+ **Sequence-agnostic whole-liver segmentation on abdominal MRI.**
17
+
18
+ LiverSegMRI is a five-fold nnU-Net v2 ensemble (`3d_fullres`, `nnUNetTrainerNoMirroring`) that segments the whole liver on a single abdominal MRI sequence. It needs no sequence labels and handles:
19
+
20
+ - T1-weighted in- and opposed-phase
21
+ - pre-contrast and dynamic post-contrast (early and late arterial, portal venous, transitional or delayed)
22
+ - hepatobiliary phase
23
+ - T2-weighted
24
+ - diffusion-weighted and ADC
25
+
26
+ | Resource | Link |
27
+ |---|---|
28
+ | Paper | *Radiology: Artificial Intelligence* 2026, doi: [10.1148/ryai.XXXXXXX](https://doi.org/10.1148/ryai.XXXXXXX) |
29
+ | Code | [github.com/TaouliLab/LiverSegMRI](https://github.com/TaouliLab/LiverSegMRI) |
30
+ | Archive | [doi.org/10.5281/zenodo.XXXXXXX](https://doi.org/10.5281/zenodo.XXXXXXX) |
31
+
32
+ ## Usage
33
+
34
+ ```bash
35
+ pip install git+https://github.com/TaouliLab/LiverSegMRI.git
36
+ liversegmri predict -i /path/to/images -o /path/to/output
37
+ ```
38
+
39
+ ```python
40
+ from liversegmri.inference import LiverSegMRI
41
+
42
+ model = LiverSegMRI(device="cuda") # downloads this repository on first use
43
+ model.predict(["/path/to/pvp.nii.gz"], output_dir="/path/to/output")
44
+ ```
45
+
46
+ The model folder can also be used directly with nnU-Net v2:
47
+
48
+ ```python
49
+ from huggingface_hub import snapshot_download
50
+ from nnunetv2.inference.predict_from_raw_data import nnUNetPredictor
51
+
52
+ model_dir = snapshot_download("TaouliLab/LiverSegMRI")
53
+ predictor = nnUNetPredictor(tile_step_size=0.5, use_gaussian=True, use_mirroring=False)
54
+ predictor.initialize_from_trained_model_folder(model_dir, use_folds=(0, 1, 2, 3, 4), checkpoint_name="checkpoint_final.pth")
55
+ ```
56
+
57
+ The study applied largest-connected-component filtering to the output, as the LiverSegMRI package does by default.
58
+
59
+ ## Files
60
+
61
+ | File | Description |
62
+ |---|---|
63
+ | `dataset.json` | nnU-Net dataset description (single channel; labels background = 0, liver = 1) |
64
+ | `plans.json` | nnU-Net plans (`3d_fullres`, 1 × 1 × 1 mm target spacing, patch size 96 × 160 × 160) |
65
+ | `fold_{0..4}/checkpoint_final.pth` | Final-epoch checkpoint of each cross-validation fold, exactly as trained |
66
+
67
+ ## Training data
68
+
69
+ - **Training cohort:** 10,494 MRI sequences from 1,058 patients.
70
+ - **Sources:**
71
+ - Icahn School of Medicine at Mount Sinai (USA): chronic liver disease with LI-RADS 3/4 observations; intrahepatic cholangiocarcinoma.
72
+ - Seoul National University Hospital (South Korea): intrahepatic cholangiocarcinoma.
73
+ - The public LLD-MMRI dataset (China).
74
+ - **Reference masks:** drawn manually from scratch on every sequence by radiologists in 3D Slicer, with adjudication by an abdominal radiologist.
75
+ - **Segmentation convention:** the liver label includes the parenchyma and all intrahepatic lesions; the gallbladder and major vessels are excluded.
76
+
77
+ ## Training procedure
78
+
79
+ | Setting | Value |
80
+ |---|---|
81
+ | Framework | nnU-Net v2, `3d_fullres`, `nnUNetTrainerNoMirroring` |
82
+ | Schedule | 1,000 epochs × 250 iterations per fold, batch size 2 |
83
+ | Optimization | Initial learning rate 0.01 with polynomial decay; weight decay 3 × 10⁻⁵; deep supervision |
84
+ | Hardware | One NVIDIA A100 (40 GB) per fold, about 45 hours per fold |
85
+
86
+ ## Evaluation
87
+
88
+ Patient-wise mean Dice:
89
+
90
+ | Test set | Patients | LiverSegMRI | TotalSegmentator MRI | MRAnnotator |
91
+ |---|---:|---:|---:|---:|
92
+ | External: Duke Liver Dataset and CirrMRI600+ | 487 | **0.944** | 0.886 | 0.889 |
93
+ | Internal: held out | 188 | **0.984** | 0.728 | 0.760 |
94
+ | Development: out-of-fold | 1,058 | 0.983 | — | — |
95
+
96
+ External test set, LiverSegMRI vs comparators:
97
+
98
+ | Metric | LiverSegMRI | TotalSegmentator MRI | MRAnnotator |
99
+ |---|---:|---:|---:|
100
+ | HD95 (mm) | 6.6 | 15.3 | 27.0 |
101
+ | Volume error (%) | 5.0 | 13.9 | 12.5 |
102
+
103
+ - **By field strength:** in Mount Sinai patients with DICOM metadata, performance was similar at 1.5 T and 3.0 T.
104
+ - **Comparisons with the other models:** all paired comparisons were significant with Holm-adjusted P < .001.
105
+
106
+ See the paper for sequence-wise, subgroup, failure, and label-matched analyses.
107
+
108
+ ## Intended use and limitations
109
+
110
+ - **Intended use:** research on whole-liver segmentation of axial abdominal MRI in adults. Examples are volumetry, quantitative MRI, and localization for downstream analyses.
111
+ - **Not a medical device.** Outputs must be reviewed and must not be used for clinical decision-making.
112
+ - **Performance is lower on external data** (Dice 0.944) than on internal data. It may degrade further with severe artifacts, low-field or nonstandard protocols, non-axial acquisitions, or pediatric patients.
113
+ - **Gross failures (Dice < 0.50) were rare:** 2 of 1,047 external sequences, both T2-weighted images of cirrhotic livers with ascites.
114
+ - **Whole liver only:** no segmental, vascular, or lesion labels.
115
+
116
+ ## Citation
117
+
118
+ ```bibtex
119
+ @article{yuce2026liversegmri,
120
+ title = {LiverSegMRI: Development and External Validation of a Deep Learning Model for Automated Liver Segmentation across Multiparametric MRI Sequences and Comparison with Generalist Segmentation Models},
121
+ author = {Yuce, Murat and Tordjman, Mickael and Meribout, Anis and Ozkaya, Efe and Lee, Jung-Oh and Lee, Jeong Min and Akinci D'Antonoli, Tugba and Wasserthal, Jakob and Mei, Xueyan and Taouli, Bachir},
122
+ journal = {Radiology: Artificial Intelligence},
123
+ year = {2026},
124
+ doi = {10.1148/ryai.XXXXXXX}
125
+ }
126
+ ```
127
+
128
+ ## License
129
+
130
+ [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/): non-commercial research use with attribution.
dataset.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "channel_names": {
3
+ "0": "LiverSegmentation"
4
+ },
5
+ "labels": {
6
+ "background": 0,
7
+ "liver": 1
8
+ },
9
+ "numTraining": 10494,
10
+ "file_ending": ".nrrd",
11
+ "overwrite_image_reader_writer": "SimpleITKIO"
12
+ }
fold_0/checkpoint_final.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be4ce40f800a077d84594943a0a167afa2c97bb30ceb09f9f20aac644c3e4e27
3
+ size 246741213
fold_1/checkpoint_final.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4104f4f00fe6d1d42129cf5f1fd798f94bc315edc26742dfaef3ef89f46ba11d
3
+ size 246741149
fold_2/checkpoint_final.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:160ed3c31b0a84d02e2296f8467620e0207fe812bda517ee764ef6a263f828d7
3
+ size 246741149
fold_3/checkpoint_final.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4c33904c832195bf82140bc82635040b2f97c1803557f021ba95a7907663231c
3
+ size 246741149
fold_4/checkpoint_final.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:938ad83f28865aa63451c05eb7226de8163be852bf5bab5f61134368e1c9744d
3
+ size 246741213
plans.json ADDED
@@ -0,0 +1,521 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "Dataset001_LiverSegmentation",
3
+ "plans_name": "nnUNetPlans",
4
+ "original_median_spacing_after_transp": [
5
+ 3.4999969005584717,
6
+ 0.9943181872367859,
7
+ 0.9943181872367859
8
+ ],
9
+ "original_median_shape_after_transp": [
10
+ 64,
11
+ 310,
12
+ 352
13
+ ],
14
+ "image_reader_writer": "SimpleITKIO",
15
+ "transpose_forward": [
16
+ 0,
17
+ 1,
18
+ 2
19
+ ],
20
+ "transpose_backward": [
21
+ 0,
22
+ 1,
23
+ 2
24
+ ],
25
+ "configurations": {
26
+ "2d": {
27
+ "data_identifier": "nnUNetPlans_2d",
28
+ "preprocessor_name": "DefaultPreprocessor",
29
+ "batch_size": 22,
30
+ "patch_size": [
31
+ 384,
32
+ 384
33
+ ],
34
+ "median_image_size_in_voxels": [
35
+ 324.0,
36
+ 379.0
37
+ ],
38
+ "spacing": [
39
+ 1.0,
40
+ 1.0
41
+ ],
42
+ "normalization_schemes": [
43
+ "ZScoreNormalization"
44
+ ],
45
+ "use_mask_for_norm": [
46
+ false
47
+ ],
48
+ "resampling_fn_data": "resample_data_or_seg_to_shape",
49
+ "resampling_fn_seg": "resample_data_or_seg_to_shape",
50
+ "resampling_fn_data_kwargs": {
51
+ "is_seg": false,
52
+ "order": 3,
53
+ "order_z": 0,
54
+ "force_separate_z": null
55
+ },
56
+ "resampling_fn_seg_kwargs": {
57
+ "is_seg": true,
58
+ "order": 1,
59
+ "order_z": 0,
60
+ "force_separate_z": null
61
+ },
62
+ "resampling_fn_probabilities": "resample_data_or_seg_to_shape",
63
+ "resampling_fn_probabilities_kwargs": {
64
+ "is_seg": false,
65
+ "order": 1,
66
+ "order_z": 0,
67
+ "force_separate_z": null
68
+ },
69
+ "architecture": {
70
+ "network_class_name": "dynamic_network_architectures.architectures.unet.PlainConvUNet",
71
+ "arch_kwargs": {
72
+ "n_stages": 7,
73
+ "features_per_stage": [
74
+ 32,
75
+ 64,
76
+ 128,
77
+ 256,
78
+ 512,
79
+ 512,
80
+ 512
81
+ ],
82
+ "conv_op": "torch.nn.modules.conv.Conv2d",
83
+ "kernel_sizes": [
84
+ [
85
+ 3,
86
+ 3
87
+ ],
88
+ [
89
+ 3,
90
+ 3
91
+ ],
92
+ [
93
+ 3,
94
+ 3
95
+ ],
96
+ [
97
+ 3,
98
+ 3
99
+ ],
100
+ [
101
+ 3,
102
+ 3
103
+ ],
104
+ [
105
+ 3,
106
+ 3
107
+ ],
108
+ [
109
+ 3,
110
+ 3
111
+ ]
112
+ ],
113
+ "strides": [
114
+ [
115
+ 1,
116
+ 1
117
+ ],
118
+ [
119
+ 2,
120
+ 2
121
+ ],
122
+ [
123
+ 2,
124
+ 2
125
+ ],
126
+ [
127
+ 2,
128
+ 2
129
+ ],
130
+ [
131
+ 2,
132
+ 2
133
+ ],
134
+ [
135
+ 2,
136
+ 2
137
+ ],
138
+ [
139
+ 2,
140
+ 2
141
+ ]
142
+ ],
143
+ "n_conv_per_stage": [
144
+ 2,
145
+ 2,
146
+ 2,
147
+ 2,
148
+ 2,
149
+ 2,
150
+ 2
151
+ ],
152
+ "n_conv_per_stage_decoder": [
153
+ 2,
154
+ 2,
155
+ 2,
156
+ 2,
157
+ 2,
158
+ 2
159
+ ],
160
+ "conv_bias": true,
161
+ "norm_op": "torch.nn.modules.instancenorm.InstanceNorm2d",
162
+ "norm_op_kwargs": {
163
+ "eps": 1e-05,
164
+ "affine": true
165
+ },
166
+ "dropout_op": null,
167
+ "dropout_op_kwargs": null,
168
+ "nonlin": "torch.nn.LeakyReLU",
169
+ "nonlin_kwargs": {
170
+ "inplace": true
171
+ }
172
+ },
173
+ "_kw_requires_import": [
174
+ "conv_op",
175
+ "norm_op",
176
+ "dropout_op",
177
+ "nonlin"
178
+ ]
179
+ },
180
+ "batch_dice": true
181
+ },
182
+ "3d_lowres": {
183
+ "data_identifier": "nnUNetPlans_3d_lowres",
184
+ "preprocessor_name": "DefaultPreprocessor",
185
+ "batch_size": 2,
186
+ "patch_size": [
187
+ 96,
188
+ 160,
189
+ 160
190
+ ],
191
+ "median_image_size_in_voxels": [
192
+ 154,
193
+ 227,
194
+ 266
195
+ ],
196
+ "spacing": [
197
+ 1.4257608868461793,
198
+ 1.4257608868461793,
199
+ 1.4257608868461793
200
+ ],
201
+ "normalization_schemes": [
202
+ "ZScoreNormalization"
203
+ ],
204
+ "use_mask_for_norm": [
205
+ false
206
+ ],
207
+ "resampling_fn_data": "resample_data_or_seg_to_shape",
208
+ "resampling_fn_seg": "resample_data_or_seg_to_shape",
209
+ "resampling_fn_data_kwargs": {
210
+ "is_seg": false,
211
+ "order": 3,
212
+ "order_z": 0,
213
+ "force_separate_z": null
214
+ },
215
+ "resampling_fn_seg_kwargs": {
216
+ "is_seg": true,
217
+ "order": 1,
218
+ "order_z": 0,
219
+ "force_separate_z": null
220
+ },
221
+ "resampling_fn_probabilities": "resample_data_or_seg_to_shape",
222
+ "resampling_fn_probabilities_kwargs": {
223
+ "is_seg": false,
224
+ "order": 1,
225
+ "order_z": 0,
226
+ "force_separate_z": null
227
+ },
228
+ "architecture": {
229
+ "network_class_name": "dynamic_network_architectures.architectures.unet.PlainConvUNet",
230
+ "arch_kwargs": {
231
+ "n_stages": 6,
232
+ "features_per_stage": [
233
+ 32,
234
+ 64,
235
+ 128,
236
+ 256,
237
+ 320,
238
+ 320
239
+ ],
240
+ "conv_op": "torch.nn.modules.conv.Conv3d",
241
+ "kernel_sizes": [
242
+ [
243
+ 3,
244
+ 3,
245
+ 3
246
+ ],
247
+ [
248
+ 3,
249
+ 3,
250
+ 3
251
+ ],
252
+ [
253
+ 3,
254
+ 3,
255
+ 3
256
+ ],
257
+ [
258
+ 3,
259
+ 3,
260
+ 3
261
+ ],
262
+ [
263
+ 3,
264
+ 3,
265
+ 3
266
+ ],
267
+ [
268
+ 3,
269
+ 3,
270
+ 3
271
+ ]
272
+ ],
273
+ "strides": [
274
+ [
275
+ 1,
276
+ 1,
277
+ 1
278
+ ],
279
+ [
280
+ 2,
281
+ 2,
282
+ 2
283
+ ],
284
+ [
285
+ 2,
286
+ 2,
287
+ 2
288
+ ],
289
+ [
290
+ 2,
291
+ 2,
292
+ 2
293
+ ],
294
+ [
295
+ 2,
296
+ 2,
297
+ 2
298
+ ],
299
+ [
300
+ 1,
301
+ 2,
302
+ 2
303
+ ]
304
+ ],
305
+ "n_conv_per_stage": [
306
+ 2,
307
+ 2,
308
+ 2,
309
+ 2,
310
+ 2,
311
+ 2
312
+ ],
313
+ "n_conv_per_stage_decoder": [
314
+ 2,
315
+ 2,
316
+ 2,
317
+ 2,
318
+ 2
319
+ ],
320
+ "conv_bias": true,
321
+ "norm_op": "torch.nn.modules.instancenorm.InstanceNorm3d",
322
+ "norm_op_kwargs": {
323
+ "eps": 1e-05,
324
+ "affine": true
325
+ },
326
+ "dropout_op": null,
327
+ "dropout_op_kwargs": null,
328
+ "nonlin": "torch.nn.LeakyReLU",
329
+ "nonlin_kwargs": {
330
+ "inplace": true
331
+ }
332
+ },
333
+ "_kw_requires_import": [
334
+ "conv_op",
335
+ "norm_op",
336
+ "dropout_op",
337
+ "nonlin"
338
+ ]
339
+ },
340
+ "batch_dice": false,
341
+ "next_stage": "3d_cascade_fullres"
342
+ },
343
+ "3d_fullres": {
344
+ "data_identifier": "nnUNetPlans_3d_fullres",
345
+ "preprocessor_name": "DefaultPreprocessor",
346
+ "batch_size": 2,
347
+ "patch_size": [
348
+ 96,
349
+ 160,
350
+ 160
351
+ ],
352
+ "median_image_size_in_voxels": [
353
+ 220.0,
354
+ 324.0,
355
+ 379.0
356
+ ],
357
+ "spacing": [
358
+ 1.0,
359
+ 1.0,
360
+ 1.0
361
+ ],
362
+ "normalization_schemes": [
363
+ "ZScoreNormalization"
364
+ ],
365
+ "use_mask_for_norm": [
366
+ false
367
+ ],
368
+ "resampling_fn_data": "resample_data_or_seg_to_shape",
369
+ "resampling_fn_seg": "resample_data_or_seg_to_shape",
370
+ "resampling_fn_data_kwargs": {
371
+ "is_seg": false,
372
+ "order": 3,
373
+ "order_z": 0,
374
+ "force_separate_z": null
375
+ },
376
+ "resampling_fn_seg_kwargs": {
377
+ "is_seg": true,
378
+ "order": 1,
379
+ "order_z": 0,
380
+ "force_separate_z": null
381
+ },
382
+ "resampling_fn_probabilities": "resample_data_or_seg_to_shape",
383
+ "resampling_fn_probabilities_kwargs": {
384
+ "is_seg": false,
385
+ "order": 1,
386
+ "order_z": 0,
387
+ "force_separate_z": null
388
+ },
389
+ "architecture": {
390
+ "network_class_name": "dynamic_network_architectures.architectures.unet.PlainConvUNet",
391
+ "arch_kwargs": {
392
+ "n_stages": 6,
393
+ "features_per_stage": [
394
+ 32,
395
+ 64,
396
+ 128,
397
+ 256,
398
+ 320,
399
+ 320
400
+ ],
401
+ "conv_op": "torch.nn.modules.conv.Conv3d",
402
+ "kernel_sizes": [
403
+ [
404
+ 3,
405
+ 3,
406
+ 3
407
+ ],
408
+ [
409
+ 3,
410
+ 3,
411
+ 3
412
+ ],
413
+ [
414
+ 3,
415
+ 3,
416
+ 3
417
+ ],
418
+ [
419
+ 3,
420
+ 3,
421
+ 3
422
+ ],
423
+ [
424
+ 3,
425
+ 3,
426
+ 3
427
+ ],
428
+ [
429
+ 3,
430
+ 3,
431
+ 3
432
+ ]
433
+ ],
434
+ "strides": [
435
+ [
436
+ 1,
437
+ 1,
438
+ 1
439
+ ],
440
+ [
441
+ 2,
442
+ 2,
443
+ 2
444
+ ],
445
+ [
446
+ 2,
447
+ 2,
448
+ 2
449
+ ],
450
+ [
451
+ 2,
452
+ 2,
453
+ 2
454
+ ],
455
+ [
456
+ 2,
457
+ 2,
458
+ 2
459
+ ],
460
+ [
461
+ 1,
462
+ 2,
463
+ 2
464
+ ]
465
+ ],
466
+ "n_conv_per_stage": [
467
+ 2,
468
+ 2,
469
+ 2,
470
+ 2,
471
+ 2,
472
+ 2
473
+ ],
474
+ "n_conv_per_stage_decoder": [
475
+ 2,
476
+ 2,
477
+ 2,
478
+ 2,
479
+ 2
480
+ ],
481
+ "conv_bias": true,
482
+ "norm_op": "torch.nn.modules.instancenorm.InstanceNorm3d",
483
+ "norm_op_kwargs": {
484
+ "eps": 1e-05,
485
+ "affine": true
486
+ },
487
+ "dropout_op": null,
488
+ "dropout_op_kwargs": null,
489
+ "nonlin": "torch.nn.LeakyReLU",
490
+ "nonlin_kwargs": {
491
+ "inplace": true
492
+ }
493
+ },
494
+ "_kw_requires_import": [
495
+ "conv_op",
496
+ "norm_op",
497
+ "dropout_op",
498
+ "nonlin"
499
+ ]
500
+ },
501
+ "batch_dice": true
502
+ },
503
+ "3d_cascade_fullres": {
504
+ "inherits_from": "3d_fullres",
505
+ "previous_stage": "3d_lowres"
506
+ }
507
+ },
508
+ "experiment_planner_used": "ExperimentPlanner",
509
+ "label_manager": "LabelManager",
510
+ "foreground_intensity_properties_per_channel": {
511
+ "0": {
512
+ "max": 32767.0,
513
+ "mean": 362.50750732421875,
514
+ "median": 167.0,
515
+ "min": -371.0,
516
+ "percentile_00_5": 3.0,
517
+ "percentile_99_5": 3696.0,
518
+ "std": 575.3319702148438
519
+ }
520
+ }
521
+ }