-
Notifications
You must be signed in to change notification settings - Fork 0
/
metric.py
476 lines (373 loc) · 16.4 KB
/
metric.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
import numpy as np
import pandas as pd
import torch
import cv2
import omegaconf
from scipy.spatial.distance import directed_hausdorff
import SimpleITK as sitk
def measure_metrics(metric_data, metric_groups, return_tensor=False):
"""
Wrapper function for calculating all metrics
Args:
metric_data: (dict) data used for calculation of metrics, could be Tensor or Numpy Array
metric_groups: (list of strings) name of metric groups
return_tensor: (bool) return Torch Tensor if True
Returns:
metrics_results: (dict) {metric_name: metric_value}
"""
# cast Tensor to Numpy Array if needed
for k, x in metric_data.items():
if isinstance(x, torch.Tensor):
metric_data[k] = x.cpu().numpy()
# keys must match metric_groups and params.metric_groups
# (using groups to share pre-scripts)
metric_group_fns = {'disp_metrics': measure_disp_metrics,
'image_metrics': measure_image_metrics,
'seg_metrics': measure_seg_metrics,}
metric_results = dict()
for group in metric_groups:
metric_results.update(metric_group_fns[group](metric_data))
# cast Numpy arrary to Tensor if needed
if return_tensor:
for k, x in metric_results.items():
metric_results[k] = torch.tensor(x)
return metric_results
"""
Functions calculating groups of metrics
"""
def measure_disp_metrics(metric_data):
"""
Calculate DVF-related metrics.
If roi_mask is given, the disp is masked and only evaluate in the bounding box of the mask.
Args:
metric_data: (dict)
Returns:
metric_results: (dict)
"""
# new object to avoid changing data in metric_data
disp_pred = metric_data['disp_pred']
if 'disp_gt' in metric_data.keys():
disp_gt = metric_data['disp_gt']
# mask the disp with roi mask if given
if 'roi_mask' in metric_data.keys():
roi_mask = metric_data['roi_mask'] # (N, 1, *(sizes))
# find roi mask bbox mask
mask_bbox, mask_bbox_mask = bbox_from_mask(roi_mask[:, 0, ...])
# mask and bbox crop dvf gt and pred by roi_mask
disp_pred = disp_pred * roi_mask
disp_pred = bbox_crop(disp_pred, mask_bbox)
if 'disp_gt' in metric_data.keys():
disp_gt = disp_gt * roi_mask
disp_gt = bbox_crop(disp_gt, mask_bbox)
# Regularity (Jacobian) metrics
folding_ratio, mag_det_jac_det = calculate_jacobian_metrics(disp_pred)
disp_metric_results = dict()
disp_metric_results.update({'folding_ratio': folding_ratio,
'mag_det_jac_det': mag_det_jac_det})
# DVF accuracy metrics if ground truth is available
if 'disp_gt' in metric_data.keys():
disp_metric_results.update({'aee': calculate_aee(disp_pred, disp_gt),
'rmse_disp': calculate_rmse_disp(disp_pred, disp_gt)})
return disp_metric_results
def measure_image_metrics(metric_data):
# unpack metric data, keys must match metric_data input
img = metric_data['target']
img_pred = metric_data['target_pred'] # (N, 1, *sizes)
# crop out image by the roi mask bounding box if given
if 'roi_mask' in metric_data.keys():
roi_mask = metric_data['roi_mask']
mask_bbox, mask_bbox_mask = bbox_from_mask(roi_mask[:, 0, ...])
img = bbox_crop(img, mask_bbox)
img_pred = bbox_crop(img_pred, mask_bbox)
return {'rmse': calculate_rmse(img, img_pred)}
def measure_seg_metrics(metric_data):
""" Calculate segmentation """
seg_gt = metric_data['target_seg']
seg_pred = metric_data['warped_source_seg']
seg_gt = seg_gt[np.newaxis, ...]
seg_pred = seg_pred[np.newaxis, ...]
assert seg_gt.ndim == seg_pred.ndim
assert seg_gt.ndim in (4, 5) # (N, 1, *2D sizes) or (N, 1, *3D sizes)
results = dict()
for label_cls in np.unique(seg_gt):
# calculate DICE score for each class
if label_cls == 0:
# skip background
continue
results[f'dice_class_{label_cls}'] = calculate_dice(seg_gt, seg_pred, label_class=label_cls)
# calculate mean dice
results['mean_dice'] = np.mean([dice for k, dice in results.items()])
return results
"""
Functions calculating individual metrics
"""
def calculate_aee(x, y):
"""
Average End point Error (AEE, mean over point-wise L2 norm)
Input DVF shape: (N, dim, *(sizes))
"""
return np.sqrt(((x - y) ** 2).sum(axis=1)).mean()
def calculate_rmse_disp(x, y):
"""
RMSE of DVF (square root over mean of sum squared)
Input DVF shape: (N, dim, *(sizes))
"""
return np.sqrt(((x - y) ** 2).sum(axis=1).mean())
def calculate_rmse(x, y):
"""Standard RMSE formula, square root over mean
(https://wikimedia.org/api/rest_v1/media/math/render/svg/6d689379d70cd119e3a9ed3c8ae306cafa5d516d)
"""
return np.sqrt(((x - y) ** 2).mean())
def calculate_jacobian_metrics(disp):
"""
Calculate Jacobian related regularity metrics.
Args:
disp: (numpy.ndarray, shape (N, ndim, *sizes) Displacement field
Returns:
folding_ratio: (scalar) Folding ratio (ratio of Jacobian determinant < 0 points)
mag_grad_jac_det: (scalar) Mean magnitude of the spatial gradient of Jacobian determinant
"""
folding_ratio = []
mag_grad_jac_det = []
for n in range(disp.shape[0]):
disp_n = np.moveaxis(disp[n, ...], 0, -1) # (*sizes, ndim)
jac_det_n = calculate_jacobian_det(disp_n)
folding_ratio += [(jac_det_n < 0).sum() / np.prod(jac_det_n.shape)]
mag_grad_jac_det += [np.abs(np.gradient(jac_det_n)).mean()]
return np.mean(folding_ratio), np.mean(mag_grad_jac_det)
def calculate_jacobian_det(disp):
"""
Calculate Jacobian determinant of displacement field of one image/volume (2D/3D)
Args:
disp: (numpy.ndarray, shape (*sizes, ndim)) Displacement field
Returns:
jac_det: (numpy.adarray, shape (*sizes) Point-wise Jacobian determinant
"""
disp_img = sitk.GetImageFromArray(disp, isVector=True)
jac_det_img = sitk.DisplacementFieldJacobianDeterminant(disp_img)
jac_det = sitk.GetArrayFromImage(jac_det_img)
return jac_det
def calculate_dice(mask1, mask2, label_class=0):
"""
Dice score of a specified class between two label masks.
(classes are encoded but by label class number not one-hot )
Args:
mask1: (numpy.array, shape (N, 1, *sizes)) segmentation mask 1
mask2: (numpy.array, shape (N, 1, *sizes)) segmentation mask 2
label_class: (int or float)
Returns:
volume_dice
"""
mask1_pos = (mask1 == label_class).astype(np.float32)
mask2_pos = (mask2 == label_class).astype(np.float32)
assert mask1.ndim == mask2.ndim
axes = tuple(range(2, mask1.ndim))
pos1and2 = np.sum(mask1_pos * mask2_pos, axis=axes)
pos1 = np.sum(mask1_pos, axis=axes)
pos2 = np.sum(mask2_pos, axis=axes)
return np.mean(2 * pos1and2 / (pos1 + pos2 + 1e-7))
def contour_distances_2d(image1, image2, dx=1):
"""
Calculate contour distances between binary masks.
The region of interest must be encoded by 1
Args:
image1: 2D binary mask 1
image2: 2D binary mask 2
dx: physical size of a pixel (e.g. 1.8 (mm) for UKBB)
Returns:
mean_hausdorff_dist: Hausdorff distance (mean if input are 2D stacks) in pixels
"""
# Retrieve contours as list of the coordinates of the points for each contour
# convert to contiguous array and data type uint8 as required by the cv2 function
image1 = np.ascontiguousarray(image1, dtype=np.uint8)
image2 = np.ascontiguousarray(image2, dtype=np.uint8)
# extract contour points and stack the contour points into (N, 2)
contours1, _ = cv2.findContours(image1.astype('uint8'), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
contour1_pts = np.array(contours1[0])[:, 0, :]
for i in range(1, len(contours1)):
cont1_arr = np.array(contours1[i])[:, 0, :]
contour1_pts = np.vstack([contour1_pts, cont1_arr])
contours2, _ = cv2.findContours(image2.astype('uint8'), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
contour2_pts = np.array(contours2[0])[:, 0, :]
for i in range(1, len(contours2)):
cont2_arr = np.array(contours2[i])[:, 0, :]
contour2_pts = np.vstack([contour2_pts, cont2_arr])
# distance matrix between two point sets
dist_matrix = np.zeros((contour1_pts.shape[0], contour2_pts.shape[0]))
for i in range(contour1_pts.shape[0]):
for j in range(contour2_pts.shape[0]):
dist_matrix[i, j] = np.linalg.norm(contour1_pts[i, :] - contour2_pts[j, :])
# symmetrical mean contour distance
mean_contour_dist = 0.5 * (np.mean(np.min(dist_matrix, axis=0)) + np.mean(np.min(dist_matrix, axis=1)))
# calculate Hausdorff distance using the accelerated method
# (doesn't really save computation since pair-wise distance matrix has to be computed for MCD anyways)
hausdorff_dist = directed_hausdorff(contour1_pts, contour2_pts)[0]
return mean_contour_dist * dx, hausdorff_dist * dx
def contour_distances_stack(stack1, stack2, label_class, dx=1):
"""
Measure mean contour distance metrics between two 2D stacks
Args:
stack1: stack of binary 2D images, shape format (W, H, N)
stack2: stack of binary 2D images, shape format (W, H, N)
label_class: class of which to calculate distance
dx: physical size of a pixel (e.g. 1.8 (mm) for UKBB)
Return:
mean_mcd: mean contour distance averaged over non-empty slices
mean_hd: Hausdorff distance averaged over non-empty slices
"""
# assert the two stacks has the same number of slices
assert stack1.shape[-1] == stack2.shape[-1], 'Contour dist error: two stacks has different number of slices'
# mask by class
stack1 = (stack1 == label_class).astype('uint8')
stack2 = (stack2 == label_class).astype('uint8')
mcd_buffer = []
hd_buffer = []
for slice_idx in range(stack1.shape[-1]):
# ignore empty masks
if np.sum(stack1[:, :, slice_idx]) > 0 and np.sum(stack2[:, :, slice_idx]) > 0:
slice1 = stack1[:, :, slice_idx]
slice2 = stack2[:, :, slice_idx]
mcd, hd = contour_distances_2d(slice1, slice2, dx=dx)
mcd_buffer += [mcd]
hd_buffer += [hd]
return np.mean(mcd_buffer), np.mean(hd_buffer)
class MetricReporter(object):
"""
Collect and report values
self.collect_value() collects value in `report_data_dict`, which is structured as:
self.report_data_dict = {'value_name_A': [A1, A2, ...], ... }
self.summarise() construct the report dictionary if called, which is structured as:
self.report = {'value_name_A': {'mean': A_mean,
'std': A_std,
'list': [A1, A2, ...]}
}
"""
def __init__(self, id_list, save_dir, save_name='analysis_results'):
self.id_list = id_list
self.save_dir = save_dir
self.save_name = save_name
self.report_data_dict = {}
self.report = {}
def reset(self):
self.report_data_dict = {}
self.report = {}
def collect(self, x):
for name, value in x.items():
if name not in self.report_data_dict.keys():
self.report_data_dict[name] = []
self.report_data_dict[name].append(value)
def summarise(self):
# summarise aggregated results to form the report dict
for name in self.report_data_dict:
self.report[name] = {
'mean': np.mean(self.report_data_dict[name]),
'std': np.std(self.report_data_dict[name]),
'list': self.report_data_dict[name]
}
def save_mean_std(self):
report_mean_std = {}
for metric_name in self.report:
report_mean_std[metric_name + '_mean'] = self.report[metric_name]['mean']
report_mean_std[metric_name + '_std'] = self.report[metric_name]['std']
# save to CSV
csv_path = self.save_dir + f'/{self.save_name}.csv'
save_dict_to_csv(report_mean_std, csv_path)
def save_df(self):
# method_column = [str(model_name)] * len(self.id_list)
# df_dict = {'Method': method_column, 'ID': self.id_list}
df_dict = {'ID': self.id_list}
for metric_name in self.report:
df_dict[metric_name] = self.report[metric_name]['list']
df = pd.DataFrame(data=df_dict)
df.to_pickle(self.save_dir + f'/{self.save_name}_df.pkl')
def bbox_crop(x, bbox):
"""
Crop image by slicing using bounding box indices (2D/3D)
Args:
x: (numpy.ndarray, shape (N, ch, *dims))
bbox: (list of tuples) [*(bbox_min_index, bbox_max_index)]
Returns:
x cropped using bounding box
"""
# slice all of batch and channel
slicer = [slice(0, x.shape[0]), slice(0, x.shape[1])]
# slice image dimensions
for bb in bbox:
slicer.append(slice(*bb))
return x[tuple(slicer)]
def bbox_from_mask(mask, pad_ratio=0.2):
"""
Find a bounding box indices of a mask (with positive > 0)
The output indices can be directly used for slicing
- for 2D, find the largest bounding box out of the N masks
- for 3D, find the bounding box of the volume mask
Args:
mask: (numpy.ndarray, shape (N, H, W) or (N, H, W, D)
pad_ratio: (int or tuple) the ratio of between the mask bounding box to image boundary to pad
Return:
bbox: (list of tuples) [*(bbox_min_index, bbox_max_index)]
bbox_mask: (numpy.ndarray shape (N, mH, mW) or (N, mH, mW, mD)) binary mask of the bounding box
"""
dim = mask.ndim - 1
mask_shape = mask.shape[1:]
pad_ratio = param_ndim_setup(pad_ratio, dim)
# find non-zero locations in the mask
nonzero_indices = np.nonzero(mask > 0)
bbox = [(nonzero_indices[i + 1].min(), nonzero_indices[i + 1].max())
for i in range(dim)]
# pad pad_ratio of the minimum distance
# from mask bounding box to the image boundaries (half each side)
for i in range(dim):
if pad_ratio[i] > 1:
print(f"Invalid padding value (>1) on dimension {dim}, set to 1")
pad_ratio[i] = 1
bbox_padding = [pad_ratio[i] * min(bbox[i][0], mask_shape[i] - bbox[i][1])
for i in range(dim)]
# "padding" by modifying the bounding box indices
bbox = [(bbox[i][0] - int(bbox_padding[i]/2), bbox[i][1] + int(bbox_padding[i]/2))
for i in range(dim)]
# bbox mask
bbox_mask = np.zeros(mask.shape, dtype=np.float32)
slicer = [slice(0, mask.shape[0])] # all slices/batch
for i in range(dim):
slicer.append(slice(*bbox[i]))
bbox_mask[tuple(slicer)] = 1.0
return bbox, bbox_mask
def param_ndim_setup(param, ndim):
"""
Check dimensions of paramters and extend dimension if needed.
Args:
param: (int/float, tuple or list) check dimension match if tuple or list is given,
expand to `dim` by repeating if a single integer/float number is given.
ndim: (int) data/model dimension
Returns:
param: (tuple)
"""
if isinstance(param, (int, float)):
param = (param,) * ndim
elif isinstance(param, (tuple, list, omegaconf.listconfig.ListConfig)):
assert len(param) == ndim, \
f"Dimension ({ndim}) mismatch with data"
param = tuple(param)
else:
raise TypeError("Parameter type not int, tuple or list")
return param
def save_dict_to_csv(d, csv_path, model_name='modelX'):
for k, x in d.items():
if not isinstance(x, list):
d[k] = [x]
pd.DataFrame(d, index=[model_name]).to_csv(csv_path)
def prepare_and_measure_metrics(metric_data_raw, metric_groups, transform=None, warped_image=None,
warped_image_seg=None):
# Update metric data if needed (after registration case)
if transform is not None:
metric_data_raw.update({
'disp_pred': transform.tensor(),
'target_pred': warped_image,
'warped_source_seg': warped_image_seg.int(),
})
# Convert tensor data to numpy and detach from computation graph
metric_data = {key: value.detach().cpu().numpy() for key, value in metric_data_raw.items()}
# Measure metrics
metrics_results = measure_metrics(metric_data, metric_groups, return_tensor=False)
return metrics_results