|
a |
|
b/utils.py |
|
|
1 |
#!/usr/bin/python |
|
|
2 |
import numpy as np |
|
|
3 |
from keras.models import * |
|
|
4 |
from keras.layers import * |
|
|
5 |
from keras.optimizers import * |
|
|
6 |
from keras.callbacks import * |
|
|
7 |
from keras.losses import * |
|
|
8 |
from keras.preprocessing.image import * |
|
|
9 |
from os.path import isfile |
|
|
10 |
from tqdm import tqdm |
|
|
11 |
import random |
|
|
12 |
from glob import glob |
|
|
13 |
import skimage.io as io |
|
|
14 |
import skimage.transform as tr |
|
|
15 |
import skimage.morphology as mo |
|
|
16 |
import SimpleITK as sitk |
|
|
17 |
from pushover import Client |
|
|
18 |
import matplotlib.pyplot as plt |
|
|
19 |
|
|
|
20 |
# img helper functions |
|
|
21 |
|
|
|
22 |
def print_info(x): |
|
|
23 |
print(str(x.shape) + ' - Min: ' + str(x.min()) + ' - Mean: ' + str(x.mean()) + ' - Max: ' + str(x.max())) |
|
|
24 |
|
|
|
25 |
def show_samples(x, y, num): |
|
|
26 |
two_d = True if len(x.shape) == 4 else False |
|
|
27 |
rnd = np.random.permutation(len(x)) |
|
|
28 |
for i in range(0, num, 2): |
|
|
29 |
plt.figure(figsize=(15, 5)) |
|
|
30 |
for j in range(2): |
|
|
31 |
plt.subplot(1,4,1+j*2) |
|
|
32 |
img = x[rnd[i+j], ..., 0] if two_d else x[rnd[i], 8+8*j, ..., 0] |
|
|
33 |
plt.axis('off') |
|
|
34 |
plt.imshow(img.astype('float32')) |
|
|
35 |
plt.subplot(1,4,2+j*2) |
|
|
36 |
if y[rnd[i]].shape[-1] == 1: |
|
|
37 |
img = y[rnd[i+j], ..., 0] if two_d else y[rnd[i], 8+8*j, ..., 0] |
|
|
38 |
else: |
|
|
39 |
img = y[rnd[i+j]] if two_d else y[rnd[i], 8+8*j] |
|
|
40 |
plt.axis('off') |
|
|
41 |
plt.imshow(img.astype('float32')) |
|
|
42 |
plt.show() |
|
|
43 |
|
|
|
44 |
def show_samples_2d(x, num, titles=None, axis_off=True, size=(20,20)): |
|
|
45 |
assert(len(x) >= 1) |
|
|
46 |
if titles: |
|
|
47 |
assert(len(titles) == len(x)) |
|
|
48 |
rnd = np.random.permutation(len(x[0])) |
|
|
49 |
for row in range(num): |
|
|
50 |
plt.figure(figsize=size) |
|
|
51 |
for col in range(len(x)): |
|
|
52 |
plt.subplot(1,len(x), col+1) |
|
|
53 |
img = x[col][rnd[row], ..., 0] if x[col][rnd[row]].shape[-1] == 1 else x[col][rnd[row]] |
|
|
54 |
if axis_off: |
|
|
55 |
plt.axis('off') |
|
|
56 |
if titles: |
|
|
57 |
plt.title(titles[col]) |
|
|
58 |
plt.imshow(img.astype('float32'), cmap='gray') |
|
|
59 |
plt.show() |
|
|
60 |
|
|
|
61 |
def shuffle(x, y): |
|
|
62 |
perm = np.random.permutation(len(x)) |
|
|
63 |
x = x[perm] |
|
|
64 |
y = y[perm] |
|
|
65 |
return x, y |
|
|
66 |
|
|
|
67 |
def split(x, y, tr_size): |
|
|
68 |
tr_size = int(len(x) * tr_size) |
|
|
69 |
x_tr = x[:tr_size] |
|
|
70 |
y_tr = y[:tr_size] |
|
|
71 |
x_te = x[tr_size:] |
|
|
72 |
y_te = y[tr_size:] |
|
|
73 |
return x_tr, y_tr, x_te, y_te |
|
|
74 |
|
|
|
75 |
def augment(x, y, h_shift=[], v_flip=False, h_flip=False, rot90=False, edge_mode='minimum'): |
|
|
76 |
assert(len(x.shape) == 4) |
|
|
77 |
seg = False if len(y.shape) <= 2 else True |
|
|
78 |
if h_shift and h_shift != 0 and len(h_shift) != 0: |
|
|
79 |
tmp_x, tmp_y = [], [] |
|
|
80 |
for shft in h_shift: |
|
|
81 |
if shft > 0: |
|
|
82 |
tmp = np.lib.pad(x[:, :, :-shft], ((0,0), (0,0), (shft,0), (0,0)), edge_mode) |
|
|
83 |
tmp_x.append(tmp) |
|
|
84 |
if seg: |
|
|
85 |
tmp = np.lib.pad(y[:, :, :-shft], ((0,0), (0,0), (shft,0), (0,0)), edge_mode) |
|
|
86 |
else: |
|
|
87 |
tmp = y |
|
|
88 |
tmp_y.append(tmp) |
|
|
89 |
else: |
|
|
90 |
tmp = np.lib.pad(x[:, :, -shft:], ((0,0), (0,0), (0,-shft), (0,0)), edge_mode) |
|
|
91 |
tmp_x.append(tmp) |
|
|
92 |
if seg: |
|
|
93 |
tmp = np.lib.pad(y[:, :, -shft:], ((0,0), (0,0), (0,-shft), (0,0)), edge_mode) |
|
|
94 |
else: |
|
|
95 |
tmp = y |
|
|
96 |
tmp_y.append(tmp) |
|
|
97 |
x = np.concatenate((x, np.concatenate(tmp_x))) |
|
|
98 |
y = np.concatenate((y, np.concatenate(tmp_y))) |
|
|
99 |
if v_flip: |
|
|
100 |
tmp = np.flip(x, axis=1) |
|
|
101 |
x = np.concatenate((x, tmp)) |
|
|
102 |
if seg: |
|
|
103 |
tmp = np.flip(y, axis=1) |
|
|
104 |
y = np.concatenate((y, tmp)) |
|
|
105 |
else: |
|
|
106 |
y = np.concatenate((y, y)) |
|
|
107 |
if h_flip: |
|
|
108 |
tmp = np.flip(x, axis=2) |
|
|
109 |
x = np.concatenate((x, tmp)) |
|
|
110 |
if seg: |
|
|
111 |
tmp = np.flip(y, axis=2) |
|
|
112 |
y = np.concatenate((y, tmp)) |
|
|
113 |
else: |
|
|
114 |
y = np.concatenate((y, y)) |
|
|
115 |
if rot90: |
|
|
116 |
tmp = np.rot90(x, axes=(1,2)) |
|
|
117 |
x = np.concatenate((x, tmp)) |
|
|
118 |
if seg: |
|
|
119 |
tmp = np.rot90(y, axes=(1,2)) |
|
|
120 |
y = np.concatenate((y, tmp)) |
|
|
121 |
else: |
|
|
122 |
y = np.concatenate((y, y)) |
|
|
123 |
return x, y |
|
|
124 |
|
|
|
125 |
def resize_3d(img, size): |
|
|
126 |
img2 = np.zeros((img.shape[0], size[0], size[1], img.shape[-1])) |
|
|
127 |
for i in range(img.shape[0]): |
|
|
128 |
img2[i] = tr.resize(img[i], (size[0], size[1]), mode='constant', preserve_range=True) |
|
|
129 |
return img2 |
|
|
130 |
|
|
|
131 |
def to_2d(x): |
|
|
132 |
assert len(x.shape) == 5 # Shape: (#, Z, Y, X, C) |
|
|
133 |
return np.reshape(x, (x.shape[0]*x.shape[1], x.shape[2], x.shape[3], x.shape[4])) |
|
|
134 |
|
|
|
135 |
def to_3d(imgs, z): |
|
|
136 |
assert len(imgs.shape) == 4 # Shape: (#, Y, X, C) |
|
|
137 |
return np.reshape(imgs, (imgs.shape[0] / z, z, imgs.shape[1], imgs.shape[2], imgs.shape[3])) |
|
|
138 |
|
|
|
139 |
def get_crop_area(img, threshold=0): |
|
|
140 |
y_arr = np.where(img.sum(axis=0) > threshold)[0] |
|
|
141 |
size = y_arr[-1] - y_arr[0] + 1 |
|
|
142 |
y = y_arr[0] |
|
|
143 |
x_arr = np.where(img.sum(axis=0).sum(axis=0) > threshold)[0] |
|
|
144 |
x = (x_arr[0] + x_arr[-1]) // 2 - size // 2 |
|
|
145 |
return y, x, size |
|
|
146 |
|
|
|
147 |
def n4_bias_correction(img): |
|
|
148 |
img = sitk.GetImageFromArray(img[..., 0].astype('float32')) |
|
|
149 |
mask = sitk.OtsuThreshold(img, 0, 1, 200) |
|
|
150 |
img = sitk.N4BiasFieldCorrection(img, mask) |
|
|
151 |
return sitk.GetArrayFromImage(img)[..., np.newaxis] |
|
|
152 |
|
|
|
153 |
def handle_specials(img): |
|
|
154 |
if img.shape[0] == 26: |
|
|
155 |
img = img[1:-1] |
|
|
156 |
elif img.shape[0] == 20: |
|
|
157 |
img = np.lib.pad(img, ((2,2), (0,0), (0,0), (0,0)), 'minimum') |
|
|
158 |
return img |
|
|
159 |
|
|
|
160 |
def erode(imgs, amount=3): |
|
|
161 |
imgs = imgs.sum(axis=-1) |
|
|
162 |
for i in range(len(imgs)): |
|
|
163 |
imgs[i] = mo.erosion(imgs[i], mo.square(amount)) |
|
|
164 |
return imgs[..., np.newaxis] |
|
|
165 |
|
|
|
166 |
def add_noise(imgs, amount=3): |
|
|
167 |
imgs = imgs.sum(axis=-1) |
|
|
168 |
for i in range(len(imgs)): |
|
|
169 |
if i % 2 == 0: |
|
|
170 |
imgs[i] = mo.dilation(imgs[i], mo.square(amount)) |
|
|
171 |
else: |
|
|
172 |
imgs[i] = mo.erosion(imgs[i], mo.square(amount)) |
|
|
173 |
return imgs[..., np.newaxis] |
|
|
174 |
|
|
|
175 |
|
|
|
176 |
# Label helper functions |
|
|
177 |
|
|
|
178 |
def to_classes(y, start, end, step=1): |
|
|
179 |
age_range = end - start |
|
|
180 |
num_classes = int(round(age_range / step)) |
|
|
181 |
labels = np.zeros((len(y), num_classes)) |
|
|
182 |
idx = (y - start) / step |
|
|
183 |
for i in range(len(idx)): |
|
|
184 |
labels[i, int(idx[i])] = 1 |
|
|
185 |
return labels |
|
|
186 |
|
|
|
187 |
def y_center(img, smooth=20, crop=100): |
|
|
188 |
# Get Sum of y-axis values |
|
|
189 |
y = img.sum(axis=-1).sum(axis=-1).sum(axis=0) |
|
|
190 |
# Smooth the values and apply the crop region |
|
|
191 |
y_vec = np.convolve(y, np.ones(smooth)/smooth, mode='same')[crop:-crop] |
|
|
192 |
# 2nd derivative of min will be max - get its index |
|
|
193 |
return np.gradient(np.gradient(y_vec)).argmax() + crop |
|
|
194 |
|
|
|
195 |
def lengthen(y, factor): |
|
|
196 |
arr = [] |
|
|
197 |
for el in y: |
|
|
198 |
for i in range(factor): |
|
|
199 |
arr.append(el) |
|
|
200 |
return np.array(arr) |
|
|
201 |
|
|
|
202 |
def shorten(y, factor): |
|
|
203 |
arr = [] |
|
|
204 |
for i in range(0, len(y), factor): |
|
|
205 |
arr.append(y[i]) |
|
|
206 |
return np.array(arr) |
|
|
207 |
|
|
|
208 |
def normalize(x, mean, std): |
|
|
209 |
return (x - x.mean()) / x.std() |
|
|
210 |
|
|
|
211 |
def multilabel(img, channel): |
|
|
212 |
if channel == 1: |
|
|
213 |
img[img > 0.01] = 1 |
|
|
214 |
img[img < 0.01] = 0 |
|
|
215 |
return img |
|
|
216 |
else: |
|
|
217 |
step = img.max() // channel |
|
|
218 |
divider = img.max() * 0.99 |
|
|
219 |
img2 = np.zeros((img.shape[0], img.shape[1], img.shape[2], channel)) |
|
|
220 |
for c in range(channel): |
|
|
221 |
img2[img[..., 0] > divider, c] = 1 |
|
|
222 |
img[img[..., 0] > divider, 0] = 0 |
|
|
223 |
divider -= step |
|
|
224 |
return img2 |
|
|
225 |
|
|
|
226 |
def read_mhd(path, label=0, crop=None, size=None, bias=False, norm=False): |
|
|
227 |
img = io.imread(path, plugin='simpleitk')[..., np.newaxis].astype('float64') |
|
|
228 |
img = handle_specials(img) |
|
|
229 |
img = multilabel(img, label) if label > 0 else img |
|
|
230 |
img = img[:, crop[0]:crop[0]+crop[2], crop[1]:crop[1]+crop[2]] if crop else img |
|
|
231 |
#img = img[:, crop[0]:-2*crop[1]+crop[0], crop[1]:-1*crop[1]] if crop else img |
|
|
232 |
img = resize_3d(img, size) if size else img |
|
|
233 |
img = n4_bias_correction(img) if bias else img |
|
|
234 |
img = (img - img.mean()) / img.std() if norm else img |
|
|
235 |
return img.astype('float32') |
|
|
236 |
|
|
|
237 |
def load_data(path, label=0, size=(24,224,224), bias=False, norm=False, to2d=False): |
|
|
238 |
files = glob(path) |
|
|
239 |
x, y = [], [] |
|
|
240 |
for i in tqdm(range(len(files))): |
|
|
241 |
img = read_mhd(files[i]) |
|
|
242 |
top, left, dim = get_crop_area(img) |
|
|
243 |
img = read_mhd(files[i], label=label, crop=(top, left, dim), size=size) |
|
|
244 |
if to2d: |
|
|
245 |
for layer in img: |
|
|
246 |
y.append(layer) |
|
|
247 |
else: |
|
|
248 |
y.append(img) |
|
|
249 |
files[i] = files[i].replace('/VOI_LABEL/', '/MHD/', 1) |
|
|
250 |
files[i] = files[i].replace('_LABEL.', '_ORIG.', 1) |
|
|
251 |
img = read_mhd(files[i], crop=(top, left, dim), size=size, bias=bias, norm=norm) |
|
|
252 |
if to2d: |
|
|
253 |
for layer in img: |
|
|
254 |
x.append(layer) |
|
|
255 |
else: |
|
|
256 |
x.append(img) |
|
|
257 |
x = np.array(x) |
|
|
258 |
y = np.array(y) |
|
|
259 |
return x, y |
|
|
260 |
|
|
|
261 |
def load_data_age(files, size=None, crop=None, bias=False, norm=False, |
|
|
262 |
to2d=False, smart_crop=False): |
|
|
263 |
files = glob(files) |
|
|
264 |
x, y = [], [] |
|
|
265 |
for i in tqdm(range(len(files))): |
|
|
266 |
if crop: |
|
|
267 |
if smart_crop: |
|
|
268 |
img = read_mhd(files[i]) |
|
|
269 |
c = y_center(img) |
|
|
270 |
crop[0] = c - crop[2] // 2 |
|
|
271 |
img = read_mhd(files[i], crop=crop, size=size, bias=bias, norm=norm) |
|
|
272 |
f = files[i].split('_') |
|
|
273 |
age = int(f[3]) + int(f[4]) / 12. |
|
|
274 |
if to2d: |
|
|
275 |
for layer in img: |
|
|
276 |
x.append(layer) |
|
|
277 |
y.append(age) |
|
|
278 |
else: |
|
|
279 |
x.append(img) |
|
|
280 |
y.append(age) |
|
|
281 |
x = np.array(x) |
|
|
282 |
y = np.array(y) |
|
|
283 |
return x, y |
|
|
284 |
|
|
|
285 |
def print_weights(weight_file_path): |
|
|
286 |
""" |
|
|
287 |
Prints out the structure of HDF5 file. |
|
|
288 |
|
|
|
289 |
Args: |
|
|
290 |
weight_file_path (str) : Path to the file to analyze |
|
|
291 |
""" |
|
|
292 |
f = h5py.File(weight_file_path) |
|
|
293 |
try: |
|
|
294 |
if len(f.attrs.items()): |
|
|
295 |
print("{} contains: ".format(weight_file_path)) |
|
|
296 |
print("Root attributes:") |
|
|
297 |
for key, value in f.attrs.items(): |
|
|
298 |
print(" {}: {}".format(key, value)) |
|
|
299 |
|
|
|
300 |
if len(f.items())==0: |
|
|
301 |
return |
|
|
302 |
|
|
|
303 |
for layer, g in f.items(): |
|
|
304 |
print(" {}".format(layer)) |
|
|
305 |
print(" Attributes:") |
|
|
306 |
for key, value in g.attrs.items(): |
|
|
307 |
print(" {}: {}".format(key, value)) |
|
|
308 |
|
|
|
309 |
print(" Dataset:") |
|
|
310 |
for p_name in g.keys(): |
|
|
311 |
param = g[p_name] |
|
|
312 |
print(" {}: {}".format(p_name, param.shape)) #try only "param" |
|
|
313 |
finally: |
|
|
314 |
f.close() |
|
|
315 |
|
|
|
316 |
# Models |
|
|
317 |
|
|
|
318 |
def conv_block(m, dim, acti, bn, res, do=0): |
|
|
319 |
n = Conv2D(dim, 3, activation=acti, padding='same')(m) |
|
|
320 |
n = BatchNormalization()(n) if bn else n |
|
|
321 |
n = Dropout(do)(n) if do else n |
|
|
322 |
n = Conv2D(dim, 3, activation=acti, padding='same')(n) |
|
|
323 |
n = BatchNormalization()(n) if bn else n |
|
|
324 |
return Add()([m, n]) if res else n |
|
|
325 |
|
|
|
326 |
def level_block(m, dim, depth, inc, acti, do, bn, mp, up, res): |
|
|
327 |
if depth > 0: |
|
|
328 |
n = conv_block(m, dim, acti, bn, res) |
|
|
329 |
m = MaxPooling2D()(n) if mp else Conv2D(dim, 3, strides=2, padding='same')(n) |
|
|
330 |
m = level_block(m, int(inc*dim), depth-1, inc, acti, do, bn, mp, up, res) |
|
|
331 |
if up: |
|
|
332 |
m = UpSampling2D()(m) |
|
|
333 |
m = Conv2D(dim, 2, activation=acti, padding='same')(m) |
|
|
334 |
else: |
|
|
335 |
m = Conv2DTranspose(dim, 3, strides=2, activation=acti, padding='same')(m) |
|
|
336 |
n = Add()([n, m]) |
|
|
337 |
m = conv_block(n, dim, acti, bn, res) |
|
|
338 |
else: |
|
|
339 |
m = conv_block(m, dim, acti, bn, res, do) |
|
|
340 |
return m |
|
|
341 |
|
|
|
342 |
def UNet(img_shape, out_ch=1, start_ch=32, depth=4, inc_rate=1., activation='elu', |
|
|
343 |
dropout=0.5, batchnorm=False, maxpool=True, upconv=True, residual=False): |
|
|
344 |
i = Input(shape=img_shape) |
|
|
345 |
o = level_block(i, start_ch, depth, inc_rate, activation, dropout, batchnorm, maxpool, upconv, residual) |
|
|
346 |
o = Conv2D(out_ch, 1, activation='sigmoid')(o) |
|
|
347 |
return Model(inputs=i, outputs=o) |
|
|
348 |
|
|
|
349 |
def level_block_3d(m, dim, depth, factor, acti, dropout): |
|
|
350 |
if depth > 0: |
|
|
351 |
n = Conv3D(dim, 3, activation=acti, padding='same')(m) |
|
|
352 |
n = Dropout(dropout)(n) if dropout else n |
|
|
353 |
n = Conv3D(dim, 3, activation=acti, padding='same')(n) |
|
|
354 |
m = MaxPooling3D((1,2,2))(n) |
|
|
355 |
m = level_block_3d(m, int(factor*dim), depth-1, factor, acti, dropout) |
|
|
356 |
m = UpSampling3D((1,2,2))(m) |
|
|
357 |
m = Conv3D(dim, 2, activation=acti, padding='same')(m) |
|
|
358 |
m = Concatenate(axis=4)([n, m]) |
|
|
359 |
m = Conv3D(dim, 3, activation=acti, padding='same')(m) |
|
|
360 |
return Conv3D(dim, 3, activation=acti, padding='same')(m) |
|
|
361 |
|
|
|
362 |
def UNet_3D(img_shape, n_out=1, dim=8, depth=3, factor=1.5, acti='elu', dropout=None): |
|
|
363 |
i = Input(shape=img_shape) |
|
|
364 |
o = level_block_3d(i, dim, depth, factor, acti, dropout) |
|
|
365 |
o = Conv3D(n_out, 1, activation='sigmoid')(o) |
|
|
366 |
return Model(inputs=i, outputs=o) |
|
|
367 |
|
|
|
368 |
# Loss Functions |
|
|
369 |
|
|
|
370 |
# 2TP / (2TP + FP + FN) |
|
|
371 |
def f1(y_true, y_pred): |
|
|
372 |
y_true_f = K.flatten(y_true) |
|
|
373 |
y_pred_f = K.flatten(y_pred) |
|
|
374 |
intersection = K.sum(y_true_f * y_pred_f) |
|
|
375 |
return (2. * intersection + 1.) / (K.sum(y_true_f) + K.sum(y_pred_f) + 1.) |
|
|
376 |
|
|
|
377 |
def f1_np(y_true, y_pred): |
|
|
378 |
return (2. * (y_true * y_pred).sum() + 1.) / (y_true.sum() + y_pred.sum() + 1.) |
|
|
379 |
|
|
|
380 |
def f1_loss(y_true, y_pred): |
|
|
381 |
return 1-f1(y_true, y_pred) |
|
|
382 |
|
|
|
383 |
def f2(y_true, y_pred): |
|
|
384 |
y_true_f = K.flatten(y_true) |
|
|
385 |
y_pred_f = K.flatten(y_pred) |
|
|
386 |
intersection = K.sum(y_true_f * y_pred_f) |
|
|
387 |
return (5. * intersection + 1.) / (4. * K.sum(y_true_f) + K.sum(y_pred_f) + 1.) |
|
|
388 |
|
|
|
389 |
def f2_loss(y_true, y_pred): |
|
|
390 |
return 1-f2(y_true, y_pred) |
|
|
391 |
|
|
|
392 |
dice = f1 |
|
|
393 |
dice_loss = f1_loss |
|
|
394 |
|
|
|
395 |
def iou(y_true, y_pred): |
|
|
396 |
y_true_f = K.flatten(y_true) |
|
|
397 |
y_pred_f = K.flatten(y_pred) |
|
|
398 |
intersection = K.sum(y_true_f * y_pred_f) |
|
|
399 |
return (intersection + 1.) / (K.sum(y_true_f) + K.sum(y_pred_f) + 1. - intersection) |
|
|
400 |
|
|
|
401 |
def iou_np(y_true, y_pred): |
|
|
402 |
intersection = (y_true * y_pred).sum() |
|
|
403 |
return (intersection + 1.) / (y_true.sum() + y_pred.sum() + 1. - intersection) |
|
|
404 |
|
|
|
405 |
def iou_loss(y_true, y_pred): |
|
|
406 |
return -iou(y_true, y_pred) |
|
|
407 |
|
|
|
408 |
def precision(y_true, y_pred): |
|
|
409 |
y_true_f = K.flatten(y_true) |
|
|
410 |
y_pred_f = K.flatten(y_pred) |
|
|
411 |
intersection = K.sum(y_true_f * y_pred_f) |
|
|
412 |
return (intersection + 1.) / (K.sum(y_pred_f) + 1.) |
|
|
413 |
|
|
|
414 |
def precision_np(y_true, y_pred): |
|
|
415 |
return ((y_true * y_pred).sum() + 1.) / (y_pred.sum() + 1.) |
|
|
416 |
|
|
|
417 |
def recall(y_true, y_pred): |
|
|
418 |
y_true_f = K.flatten(y_true) |
|
|
419 |
y_pred_f = K.flatten(y_pred) |
|
|
420 |
intersection = K.sum(y_true_f * y_pred_f) |
|
|
421 |
return (intersection + 1.) / (K.sum(y_true_f) + 1.) |
|
|
422 |
|
|
|
423 |
def recall_np(y_true, y_pred): |
|
|
424 |
return ((y_true * y_pred).sum() + 1.) / (y_true.sum() + 1.) |
|
|
425 |
|
|
|
426 |
def mae_img(y_true, y_pred): |
|
|
427 |
y_true_f = K.flatten(y_true) |
|
|
428 |
y_pred_f = K.flatten(y_pred) |
|
|
429 |
return mae(y_true_f, y_pred_f) |
|
|
430 |
|
|
|
431 |
def bce_img(y_true, y_pred): |
|
|
432 |
y_true_f = K.flatten(y_true) |
|
|
433 |
y_pred_f = K.flatten(y_pred) |
|
|
434 |
return binary_crossentropy(y_true_f, y_pred_f) |
|
|
435 |
|
|
|
436 |
def f1_bce(y_true, y_pred): |
|
|
437 |
return f1_loss(y_true, y_pred) + bce_img(y_true, y_pred) |
|
|
438 |
|
|
|
439 |
# FP + FN |
|
|
440 |
def error(y_true, y_pred): |
|
|
441 |
y_true_f = K.flatten(y_true) |
|
|
442 |
y_pred_f = K.flatten(y_pred) |
|
|
443 |
return K.sum(K.abs(y_true_f - y_pred_f)) / float(224*224) |
|
|
444 |
|
|
|
445 |
def error_np(y_true, y_pred): |
|
|
446 |
return (abs(y_true - y_pred)).sum() / float(len(y_true.flatten())) |
|
|
447 |
|
|
|
448 |
# Notifications |
|
|
449 |
|
|
|
450 |
def pushover(title, message): |
|
|
451 |
user = "u96ub3t5wu1nexmgi22xjs31jeb8y6" |
|
|
452 |
api = "avfytsyktracxood45myebobtry6yd" |
|
|
453 |
client = Client(user, api_token=api) |
|
|
454 |
client.send_message(message, title=title) |
|
|
455 |
|
|
|
456 |
#from nipype.interfaces.ants import N4BiasFieldCorrection |
|
|
457 |
#correct = N4BiasFieldCorrection() |
|
|
458 |
#correct.inputs.input_image = in_file |
|
|
459 |
#correct.inputs.output_image = out_file |
|
|
460 |
#done = correct.run() |
|
|
461 |
#img done.outputs.output_image |