-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
266 lines (220 loc) · 11.1 KB
/
utils.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
import numpy as np
import torch
import torch.nn as nn
import pandas as pd
import cv2 as cv
def find_majority(mask_array):
""" Finds the per-pixel majority vote of the masks and returns the result as a torch tensor.
:param mask_array: input array containing discrete masks of multiple raters; dimensions should be (img_num, rater_num, img_size, img_size)
:type mask_array: array_like
:return: majority vote of the raters (dims: (img_num, img_size, img_size))
:rtype: torch tensor
"""
img_size = mask_array.shape[2]
rater_num = mask_array.shape[1]
img_num = mask_array.shape[0]
mask_majority = torch.zeros((img_num, img_size, img_size), dtype=torch.long)
for s in range(img_num):
print('processing {}'.format(s+1))
for i in range(img_size):
for j in range(img_size):
counts = {}
counts[mask_array[s,0,i,j]] = 1
for r in range(1,rater_num):
if mask_array[s,r,i,j] not in list(counts.keys()):
counts[mask_array[s,r,i,j]] = 1
else:
counts[mask_array[s,r,i,j]] += 1
val_list = list(counts.values())
mask_majority[s,i,j] = list(counts.keys())[np.argmax(val_list)]
mask_majority = mask_majority.numpy()
print('majority vote done!')
return mask_majority
def find_variance_map(mask_array, num_classes):
""" Finds the variance map for each mask and returns the result as a torch tensor.
:param mask_array: input array containing discrete masks of multiple raters; dimensions should be (img_num, rater_num, img_size, img_size)
:type mask_array: array_like
:param num_classes: number of classes in each image (excluding background)
:type num_classes: int
:return: variance maps (dims: (img_num, num_classes, img_size, img_size))
:rtype: torch tensor
"""
img_size = mask_array.shape[2]
rater_num = mask_array.shape[1]
img_num = mask_array.shape[0]
#generate the GT variance maps for task2(variance map prediction)
mask_1hot = np.zeros((img_num, rater_num, num_classes, img_size, img_size), dtype = np.uint8)
for s in range(img_num):
for r in range(rater_num):
mask_temp = mask_array[s,r]
for i in range(1,num_classes+1):
arr = np.full((img_size, img_size), i)
mask_temp_i = (mask_temp == arr).astype(int)
mask_1hot[s,r,i-1]= mask_temp_i
log_odds = torch.zeros((img_num, rater_num, num_classes, img_size, img_size), dtype = torch.float64)
for s in range(img_num):
for r in range(rater_num):
for c in range(num_classes):
distp = cv.distanceTransform(mask_1hot[s,r,c], cv.DIST_L2, num_classes+1)
mask_complement = 1 - mask_1hot[s,r,c]
distn = cv.distanceTransform(mask_complement, cv.DIST_L2, num_classes+1)
distn = -1*distn
dist = distp + distn
log_odd = torch.special.expit(torch.from_numpy(dist))
log_odds[s,r,c] = log_odd
var_array = torch.zeros((img_num, num_classes, img_size, img_size), dtype=torch.float64)
for s in range(img_num):
for i in range(num_classes):
mean = (log_odds[s,0,i] + log_odds[s,1,i] + log_odds[s,2,i])/rater_num
variance = ((log_odds[s,0,i] - mean)**2 + (log_odds[s,1,i] - mean)**2 + (log_odds[s,2,i] - mean)**2)/rater_num
var_array[s,i] = variance
var_array = var_array.numpy()
print('variance map done!')
return var_array
def data_loader(train_path, val_path):
""" Loads the train and validation data from corresponding pandas dataframes into lists of per-image dictionaries to be used with monai dataloaders. In each dataframe, the data should be stored subject-wise.
:param train_path: path for the train dataframe
:type train_path: string
:param val_path: path for the validation dataframe
:type val_path: string
:return: lists of per-image dictionaries (train_dict, val_dict)
"""
train = pd.read_pickle(train_path)
val = pd.read_pickle(val_path)
mri_train, var_train, maj_train = [], [], []
for i in range(len(train)):
if(type(train.loc[i, 'l3l4_MRI']) != int):
mri_train.append(train.loc[i, 'l3l4_MRI'])
var_train.append(train.loc[i, 'l3l4_var'])
maj_train.append(train.loc[i, 'l3l4_maj'])
if(type(train.loc[i, 'l4l5_MRI']) != int):
mri_train.append(train.loc[i, 'l4l5_MRI'])
var_train.append(train.loc[i, 'l4l5_var'])
maj_train.append(train.loc[i, 'l4l5_maj'])
if(type(train.loc[i, 'l5s1_MRI']) != int):
mri_train.append(train.loc[i, 'l5s1_MRI'])
var_train.append(train.loc[i, 'l5s1_var'])
maj_train.append(train.loc[i, 'l5s1_maj'])
if(type(train.loc[i, 'l4upper_MRI']) != int):
mri_train.append(train.loc[i, 'l4upper_MRI'])
var_train.append(train.loc[i, 'l4upper_var'])
maj_train.append(train.loc[i, 'l4upper_maj'])
if(type(train.loc[i, 'l5upper_MRI']) != int):
mri_train.append(train.loc[i, 'l5upper_MRI'])
var_train.append(train.loc[i, 'l5upper_var'])
maj_train.append(train.loc[i, 'l5upper_maj'])
if(type(train.loc[i, 's1_MRI']) != int):
mri_train.append(train.loc[i, 's1_MRI'])
var_train.append(train.loc[i, 's1_var'])
maj_train.append(train.loc[i, 's1_maj'])
train_dict = [{'mri': mri_train[i], 'maj': maj_train[i], 'var':var_train[i]} for i in range(len(mri_train))]
mri_val, var_val, maj_val = [], [], []
for i in range(len(val)):
if(type(val.loc[i, 'l3l4_MRI']) != int):
mri_val.append(val.loc[i, 'l3l4_MRI'])
var_val.append(val.loc[i, 'l3l4_var'])
maj_val.append(val.loc[i, 'l3l4_maj'])
if(type(val.loc[i, 'l4l5_MRI']) != int):
mri_val.append(val.loc[i, 'l4l5_MRI'])
var_val.append(val.loc[i, 'l4l5_var'])
maj_val.append(val.loc[i, 'l4l5_maj'])
if(type(val.loc[i, 'l5s1_MRI']) != int):
mri_val.append(val.loc[i, 'l5s1_MRI'])
var_val.append(val.loc[i, 'l5s1_var'])
maj_val.append(val.loc[i, 'l5s1_maj'])
if(type(val.loc[i, 'l4upper_MRI']) != int):
mri_val.append(val.loc[i, 'l4upper_MRI'])
var_val.append(val.loc[i, 'l4upper_var'])
maj_val.append(val.loc[i, 'l4upper_maj'])
if(type(val.loc[i, 'l5upper_MRI']) != int):
mri_val.append(val.loc[i, 'l5upper_MRI'])
var_val.append(val.loc[i, 'l5upper_var'])
maj_val.append(val.loc[i, 'l5upper_maj'])
if(type(val.loc[i, 's1_MRI']) != int):
mri_val.append(val.loc[i, 's1_MRI'])
var_val.append(val.loc[i, 's1_var'])
maj_val.append(val.loc[i, 's1_maj'])
val_dict = [{'mri': mri_val[i], 'maj': maj_val[i], 'var':var_val[i]} for i in range(len(mri_val))]
return train_dict, val_dict
def test_loader(test_path):
""" Loads the test data from a pandas dataframe into a list of per-image dictionaries to be used with monai dataloaders. In the dataframe, the data should be stored subject-wise.
:param test_path: path for the test dataframe
:type test_path: string
:return: a list of per-image dictionaries (test_dict)
"""
test = pd.read_pickle(test_path)
mri_test, var_test, maj_test, name_test, level_test = [], [], [], [], []
for i in range(len(test)):
if(type(test.loc[i, 'l3l4_MRI']) != int):
mri_test.append(test.loc[i, 'l3l4_MRI'])
var_test.append(test.loc[i, 'l3l4_var'])
maj_test.append(test.loc[i, 'l3l4_maj'])
name_test.append(test.loc[i, 'name'])
level_test.append(1)
if(type(test.loc[i, 'l4l5_MRI']) != int):
mri_test.append(test.loc[i, 'l4l5_MRI'])
var_test.append(test.loc[i, 'l4l5_var'])
maj_test.append(test.loc[i, 'l4l5_maj'])
name_test.append(test.loc[i, 'name'])
level_test.append(2)
if(type(test.loc[i, 'l5s1_MRI']) != int):
mri_test.append(test.loc[i, 'l5s1_MRI'])
var_test.append(test.loc[i, 'l5s1_var'])
maj_test.append(test.loc[i, 'l5s1_maj'])
name_test.append(test.loc[i, 'name'])
level_test.append(3)
if(type(test.loc[i, 'l4upper_MRI']) != int):
mri_test.append(test.loc[i, 'l4upper_MRI'])
var_test.append(test.loc[i, 'l4upper_var'])
maj_test.append(test.loc[i, 'l4upper_maj'])
name_test.append(test.loc[i, 'name'])
level_test.append(4)
if(type(test.loc[i, 'l5upper_MRI']) != int):
mri_test.append(test.loc[i, 'l5upper_MRI'])
var_test.append(test.loc[i, 'l5upper_var'])
maj_test.append(test.loc[i, 'l5upper_maj'])
name_test.append(test.loc[i, 'name'])
level_test.append(5)
if(type(test.loc[i, 's1_MRI']) != int):
mri_test.append(test.loc[i, 's1_MRI'])
var_test.append(test.loc[i, 's1_var'])
maj_test.append(test.loc[i, 's1_maj'])
name_test.append(test.loc[i, 'name'])
level_test.append(6)
test_dict = [{'mri': mri_test[i], 'maj': maj_test[i], 'var':var_test[i], 'level':level_test[i]} for i in range(len(mri_test))]
return test_dict
class DiceLoss(nn.Module):
""" Class for calculating the dice loss between a GT mask and the model outputs. The outputs can be drawn from the model either before or after applying the softmax."""
def __init__(self, n_classes):
super(DiceLoss, self).__init__()
self.n_classes = n_classes
def _one_hot_encoder(self, input_tensor):
tensor_list = []
for i in range(self.n_classes):
temp_prob = input_tensor == i # * torch.ones_like(input_tensor)
tensor_list.append(temp_prob.unsqueeze(1))
output_tensor = torch.cat(tensor_list, dim=1)
return output_tensor.float()
def _dice_loss(self, score, target):
target = target.float()
smooth = 1e-5
intersect = torch.sum(score * target)
y_sum = torch.sum(target * target)
z_sum = torch.sum(score * score)
loss = (2 * intersect + smooth) / (z_sum + y_sum + smooth)
loss = 1 - loss
return loss
def forward(self, inputs, target, weight=None, softmax=False):
if softmax:
inputs = torch.softmax(inputs, dim=1)
target = self._one_hot_encoder(target)
if weight is None:
weight = [1] * self.n_classes
assert inputs.size() == target.size(), 'predict {} & target {} shape do not match'.format(inputs.size(), target.size())
class_wise_dice = []
loss = 0.0
for i in range(0, self.n_classes):
dice = self._dice_loss(inputs[:, i], target[:, i])
class_wise_dice.append(1.0 - dice.item())
loss += dice * weight[i]
return loss / self.n_classes