-
Notifications
You must be signed in to change notification settings - Fork 126
/
sample_patches_combined.py
473 lines (447 loc) · 20.6 KB
/
sample_patches_combined.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
import numpy as np
import cv2
import pandas as pd
import os, sys, argparse
import dicom
from dm_image import read_resize_img, crop_img, add_img_margins
from dm_preprocess import DMImagePreprocessor as imprep
from scipy.misc import toimage
from sklearn.model_selection import train_test_split
#### Define some functions to use ####
def const_filename(pat, side, view, directory, itype=None, abn=None):
token_list = [pat, side, view]
if itype is not None:
token_list.insert(
0, ('Calc' if itype == 'calc' else 'Mass') + '-Training')
token_list.append(str(abn))
fn = "_".join(token_list) + ".png"
return os.path.join(directory, fn)
def crop_val(v, minv, maxv):
v = v if v >= minv else minv
v = v if v <= maxv else maxv
return v
def overlap_patch_roi(patch_center, patch_size, roi_mask,
add_val=1000, cutoff=.5):
x1,y1 = (patch_center[0] - patch_size/2,
patch_center[1] - patch_size/2)
x2,y2 = (patch_center[0] + patch_size/2,
patch_center[1] + patch_size/2)
x1 = crop_val(x1, 0, roi_mask.shape[1])
y1 = crop_val(y1, 0, roi_mask.shape[0])
x2 = crop_val(x2, 0, roi_mask.shape[1])
y2 = crop_val(y2, 0, roi_mask.shape[0])
roi_area = (roi_mask>0).sum()
roi_patch_added = roi_mask.copy()
roi_patch_added[y1:y2, x1:x2] += add_val
patch_area = (roi_patch_added>=add_val).sum()
inter_area = (roi_patch_added>add_val).sum().astype('float32')
return (inter_area/roi_area > cutoff or inter_area/patch_area > cutoff)
def create_blob_detector(roi_size=(128, 128), blob_min_area=3,
blob_min_int=.5, blob_max_int=.95, blob_th_step=10):
params = cv2.SimpleBlobDetector_Params()
params.filterByArea = True
params.minArea = blob_min_area
params.maxArea = roi_size[0]*roi_size[1]
params.filterByCircularity = False
params.filterByColor = False
params.filterByConvexity = False
params.filterByInertia = False
# blob detection only works with "uint8" images.
params.minThreshold = int(blob_min_int*255)
params.maxThreshold = int(blob_max_int*255)
params.thresholdStep = blob_th_step
ver = (cv2.__version__).split('.')
if int(ver[0]) < 3:
return cv2.SimpleBlobDetector(params)
else:
return cv2.SimpleBlobDetector_create(params)
def sample_patches(img, roi_mask, out_dir, img_id, abn, pos, patch_size=256,
pos_cutoff=.75, neg_cutoff=.35,
nb_bkg=100, nb_abn=100, start_sample_nb=0, itype='calc',
bkg_dir='background',
calc_pos_dir='calc_mal', calc_neg_dir='calc_ben',
mass_pos_dir='mass_mal', mass_neg_dir='mass_ben',
verbose=False):
if pos:
if itype == 'calc':
roi_out = os.path.join(out_dir, calc_pos_dir)
else:
roi_out = os.path.join(out_dir, mass_pos_dir)
else:
if itype == 'calc':
roi_out = os.path.join(out_dir, calc_neg_dir)
else:
roi_out = os.path.join(out_dir, mass_neg_dir)
bkg_out = os.path.join(out_dir, bkg_dir)
basename = '_'.join([img_id, str(abn)])
img = add_img_margins(img, patch_size/2)
roi_mask = add_img_margins(roi_mask, patch_size/2)
# Get ROI bounding box.
roi_mask_8u = roi_mask.astype('uint8')
ver = (cv2.__version__).split('.')
if int(ver[0]) < 3:
contours,_ = cv2.findContours(
roi_mask_8u.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
else:
_,contours,_ = cv2.findContours(
roi_mask_8u.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cont_areas = [ cv2.contourArea(cont) for cont in contours ]
idx = np.argmax(cont_areas) # find the largest contour.
rx,ry,rw,rh = cv2.boundingRect(contours[idx])
if verbose:
M = cv2.moments(contours[idx])
try:
cx = int(M['m10']/M['m00'])
cy = int(M['m01']/M['m00'])
print "ROI centroid=", (cx,cy); sys.stdout.flush()
except ZeroDivisionError:
cx = rx + int(rw/2)
cy = ry + int(rh/2)
print "ROI centroid=Unknown, use b-box center=", (cx,cy)
sys.stdout.flush()
rng = np.random.RandomState(12345)
# Sample abnormality first.
sampled_abn = 0
nb_try = 0
while sampled_abn < nb_abn:
if nb_abn > 1:
x = rng.randint(rx, rx + rw)
y = rng.randint(ry, ry + rh)
nb_try += 1
if nb_try >= 1000:
print "Nb of trials reached maximum, decrease overlap cutoff by 0.05"
sys.stdout.flush()
pos_cutoff -= .05
nb_try = 0
if pos_cutoff <= .0:
raise Exception("overlap cutoff becomes non-positive, "
"check roi mask input.")
else:
x = cx
y = cy
# import pdb; pdb.set_trace()
if nb_abn == 1 or overlap_patch_roi((x,y), patch_size, roi_mask,
cutoff=pos_cutoff):
patch = img[y - patch_size/2:y + patch_size/2,
x - patch_size/2:x + patch_size/2]
patch = patch.astype('int32')
patch_img = toimage(patch, high=patch.max(), low=patch.min(),
mode='I')
# patch = patch.reshape((patch.shape[0], patch.shape[1], 1))
filename = basename + "_%04d" % (sampled_abn) + ".png"
fullname = os.path.join(roi_out, filename)
# import pdb; pdb.set_trace()
patch_img.save(fullname)
sampled_abn += 1
nb_try = 0
if verbose:
print "sampled an abn patch at (x,y) center=", (x,y)
sys.stdout.flush()
# Sample background.
sampled_bkg = start_sample_nb
while sampled_bkg < start_sample_nb + nb_bkg:
x = rng.randint(patch_size/2, img.shape[1] - patch_size/2)
y = rng.randint(patch_size/2, img.shape[0] - patch_size/2)
if not overlap_patch_roi((x,y), patch_size, roi_mask, cutoff=neg_cutoff):
patch = img[y - patch_size/2:y + patch_size/2,
x - patch_size/2:x + patch_size/2]
patch = patch.astype('int32')
patch_img = toimage(patch, high=patch.max(), low=patch.min(),
mode='I')
filename = basename + "_%04d" % (sampled_bkg) + ".png"
fullname = os.path.join(bkg_out, filename)
patch_img.save(fullname)
sampled_bkg += 1
if verbose:
print "sampled a bkg patch at (x,y) center=", (x,y)
sys.stdout.flush()
def sample_hard_negatives(img, roi_mask, out_dir, img_id, abn,
patch_size=256, neg_cutoff=.35, nb_bkg=100,
start_sample_nb=0,
bkg_dir='background', verbose=False):
'''WARNING: the definition of hns may be problematic.
There has been study showing that the context of an ROI is also useful
for classification.
'''
bkg_out = os.path.join(out_dir, bkg_dir)
basename = '_'.join([img_id, str(abn)])
img = add_img_margins(img, patch_size/2)
roi_mask = add_img_margins(roi_mask, patch_size/2)
# Get ROI bounding box.
roi_mask_8u = roi_mask.astype('uint8')
ver = (cv2.__version__).split('.')
if int(ver[0]) < 3:
contours,_ = cv2.findContours(
roi_mask_8u.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
else:
_,contours,_ = cv2.findContours(
roi_mask_8u.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cont_areas = [ cv2.contourArea(cont) for cont in contours ]
idx = np.argmax(cont_areas) # find the largest contour.
rx,ry,rw,rh = cv2.boundingRect(contours[idx])
if verbose:
M = cv2.moments(contours[idx])
cx = int(M['m10']/M['m00'])
cy = int(M['m01']/M['m00'])
print "ROI centroid=", (cx,cy); sys.stdout.flush()
rng = np.random.RandomState(12345)
# Sample hard negative samples.
sampled_bkg = start_sample_nb
while sampled_bkg < start_sample_nb + nb_bkg:
x1,x2 = (rx - patch_size/2, rx + rw + patch_size/2)
y1,y2 = (ry - patch_size/2, ry + rh + patch_size/2)
x1 = crop_val(x1, patch_size/2, img.shape[1] - patch_size/2)
x2 = crop_val(x2, patch_size/2, img.shape[1] - patch_size/2)
y1 = crop_val(y1, patch_size/2, img.shape[0] - patch_size/2)
y2 = crop_val(y2, patch_size/2, img.shape[0] - patch_size/2)
x = rng.randint(x1, x2)
y = rng.randint(y1, y2)
if not overlap_patch_roi((x,y), patch_size, roi_mask, cutoff=neg_cutoff):
patch = img[y - patch_size/2:y + patch_size/2,
x - patch_size/2:x + patch_size/2]
patch = patch.astype('int32')
patch_img = toimage(patch, high=patch.max(), low=patch.min(),
mode='I')
filename = basename + "_%04d" % (sampled_bkg) + ".png"
fullname = os.path.join(bkg_out, filename)
patch_img.save(fullname)
sampled_bkg += 1
if verbose:
print "sampled a hns patch at (x,y) center=", (x,y)
sys.stdout.flush()
def sample_blob_negatives(img, roi_mask, out_dir, img_id, abn, blob_detector,
patch_size=256, neg_cutoff=.35, nb_bkg=100,
start_sample_nb=0,
bkg_dir='background', verbose=False):
bkg_out = os.path.join(out_dir, bkg_dir)
basename = '_'.join([img_id, str(abn)])
img = add_img_margins(img, patch_size/2)
roi_mask = add_img_margins(roi_mask, patch_size/2)
# Get ROI bounding box.
roi_mask_8u = roi_mask.astype('uint8')
ver = (cv2.__version__).split('.')
if int(ver[0]) < 3:
contours,_ = cv2.findContours(
roi_mask_8u.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
else:
_,contours,_ = cv2.findContours(
roi_mask_8u.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cont_areas = [ cv2.contourArea(cont) for cont in contours ]
idx = np.argmax(cont_areas) # find the largest contour.
rx,ry,rw,rh = cv2.boundingRect(contours[idx])
if verbose:
M = cv2.moments(contours[idx])
cx = int(M['m10']/M['m00'])
cy = int(M['m01']/M['m00'])
print "ROI centroid=", (cx,cy); sys.stdout.flush()
# Sample blob negative samples.
key_pts = blob_detector.detect((img/img.max()*255).astype('uint8'))
rng = np.random.RandomState(12345)
key_pts = rng.permutation(key_pts)
sampled_bkg = 0
for kp in key_pts:
if sampled_bkg >= nb_bkg:
break
x,y = int(kp.pt[0]), int(kp.pt[1])
if not overlap_patch_roi((x,y), patch_size, roi_mask, cutoff=neg_cutoff):
patch = img[y - patch_size/2:y + patch_size/2,
x - patch_size/2:x + patch_size/2]
patch = patch.astype('int32')
patch_img = toimage(patch, high=patch.max(), low=patch.min(),
mode='I')
filename = basename + "_%04d" % (start_sample_nb + sampled_bkg) + ".png"
fullname = os.path.join(bkg_out, filename)
patch_img.save(fullname)
if verbose:
print "sampled a blob patch at (x,y) center=", (x,y)
sys.stdout.flush()
sampled_bkg += 1
return sampled_bkg
#### End of function definition ####
def run(roi_mask_path_file, roi_mask_dir, pat_train_list_file, full_img_dir,
train_out_dir, val_out_dir,
target_height=4096, target_width=None, patch_size=256,
segment_breast=True,
nb_bkg=30, nb_abn=30, nb_hns=15,
pos_cutoff=.75, neg_cutoff=.35, val_size=.1,
bkg_dir='background', calc_pos_dir='calc_mal', calc_neg_dir='calc_ben',
mass_pos_dir='mass_mal', mass_neg_dir='mass_ben', verbose=True):
# Print info for book-keeping.
print "Pathology file=", roi_mask_path_file
print "ROI mask dir=", roi_mask_dir
print "Patient train list=", pat_train_list_file
print "Full image dir=", full_img_dir
print "Train out dir=", train_out_dir
print "Val out dir=", val_out_dir
print "==="
sys.stdout.flush()
# Read ROI mask table with pathology.
roi_mask_path_df = pd.read_csv(roi_mask_path_file, header=0)
roi_mask_path_df = roi_mask_path_df.set_index(['patient_id', 'side', 'view'])
roi_mask_path_df.sort_index(inplace=True)
# Read train set patient IDs and subset the table.
pat_train = pd.read_csv(pat_train_list_file, header=None)
pat_train = pat_train.values.ravel()
if len(pat_train) > 1:
path_df = roi_mask_path_df.loc[pat_train.tolist()]
else:
locs = roi_mask_path_df.index.get_loc(pat_train[0])
path_df = roi_mask_path_df.iloc[locs]
# Determine the labels for patients.
pat_labs = []
for pat in pat_train:
pathology = path_df.loc[pat]['pathology']
malignant = 0
for path in pathology:
if path.startswith('MALIGNANT'):
malignant = 1
break
pat_labs.append(malignant)
# Split patient list into train and val lists.
def write_pat_list(fn, pat_list):
with open(fn, 'w') as f:
for pat in pat_list:
f.write(str(pat) + "\n")
f.close()
if val_size > 0:
# import pdb; pdb.set_trace()
pat_train, pat_val, labs_train, labs_val = train_test_split(
pat_train, pat_labs, stratify=pat_labs, test_size=val_size,
random_state=12345)
if len(pat_val) > 1:
val_df = roi_mask_path_df.loc[pat_val.tolist()]
else:
locs = roi_mask_path_df.index.get_loc(pat_val[0])
val_df = roi_mask_path_df.iloc[locs]
write_pat_list(os.path.join(val_out_dir, 'pat_lst.txt'), pat_val.tolist())
if len(pat_train) > 1:
train_df = roi_mask_path_df.loc[pat_train.tolist()]
else:
locs = roi_mask_path_df.index.get_loc(pat_train[0])
train_df = roi_mask_path_df.iloc[locs]
write_pat_list(os.path.join(train_out_dir, 'pat_lst.txt'), pat_train.tolist())
# Create a blob detector.
blob_detector = create_blob_detector(roi_size=(patch_size, patch_size))
#### Define a functin to sample patches.
def do_sampling(pat_df, out_dir):
for pat,side,view in pat_df.index.unique():
full_fn = const_filename(pat, side, view, full_img_dir)
# import pdb; pdb.set_trace()
try:
if target_width is None:
full_img = read_resize_img(
full_fn, target_height=target_height)
else:
full_img = read_resize_img(
full_fn, target_size=(target_height, target_width))
img_id = '_'.join([pat, side, view])
print "ID:%s, read image of size=%s" % (img_id, full_img.shape),
if segment_breast:
full_img, bbox = imprep.segment_breast(full_img)
print "size after segmentation=%s" % (str(full_img.shape))
sys.stdout.flush()
# Read mask image(s).
abn_path = roi_mask_path_df.loc[pat].loc[side].loc[view]
if isinstance(abn_path, pd.Series):
abn_num = [abn_path['abn_num']]
pathology = [abn_path['pathology']]
itypes = [abn_path['type']]
else:
abn_num = abn_path['abn_num']
pathology = abn_path['pathology']
itypes = abn_path['type']
bkg_sampled = False
for abn, path, itype in zip(abn_num, pathology, itypes):
mask_fn = const_filename(pat, side, view, roi_mask_dir, itype, abn)
if target_width is None:
mask_img = read_resize_img(
mask_fn, target_height=target_height, gs_255=True)
else:
mask_img = read_resize_img(
mask_fn, target_size=(target_height, target_width),
gs_255=True)
if segment_breast:
mask_img = crop_img(mask_img, bbox)
# sample using mask and full image.
nb_hns_ = nb_hns if not bkg_sampled else 0
if nb_hns_ > 0:
hns_sampled = sample_blob_negatives(
full_img, mask_img, out_dir, img_id,
abn, blob_detector, patch_size, neg_cutoff,
nb_hns_, 0, bkg_dir, verbose)
else:
hns_sampled = 0
pos = path.startswith('MALIGNANT')
nb_bkg_ = nb_bkg - hns_sampled if not bkg_sampled else 0
sample_patches(full_img, mask_img, out_dir, img_id, abn, pos,
patch_size, pos_cutoff, neg_cutoff,
nb_bkg_, nb_abn, hns_sampled, itype,
bkg_dir, calc_pos_dir, calc_neg_dir,
mass_pos_dir, mass_neg_dir, verbose)
bkg_sampled = True
except AttributeError:
print "Read image error: %s" % (full_fn)
except ValueError:
print "Error sampling from ROI mask image: %s" % (mask_fn)
#####
print "Sampling for train set"
sys.stdout.flush()
do_sampling(train_df, train_out_dir)
print "Done."
#####
if val_size > 0.:
print "Sampling for val set"
sys.stdout.flush()
do_sampling(val_df, val_out_dir)
print "Done."
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Sample patches for DDSM images")
parser.add_argument("roi_mask_path_file", type=str)
parser.add_argument("roi_mask_dir", type=str)
parser.add_argument("pat_train_list_file", type=str)
parser.add_argument("full_img_dir", type=str)
parser.add_argument("train_out_dir", type=str)
parser.add_argument("val_out_dir", type=str)
parser.add_argument("--target-height", dest="target_height", type=int, default=4096)
parser.add_argument("--target-width", dest="target_width", type=int, default=None)
parser.add_argument("--no-target-width", dest="target_width", action="store_const", const=None)
parser.add_argument("--segment-breast", dest="segment_breast", action="store_true")
parser.add_argument("--no-segment-breast", dest="segment_breast", action="store_false")
parser.set_defaults(segment_breast=True)
parser.add_argument("--patch-size", dest="patch_size", type=int, default=256)
parser.add_argument("--nb-bkg", dest="nb_bkg", type=int, default=30)
parser.add_argument("--nb-abn", dest="nb_abn", type=int, default=30)
parser.add_argument("--nb-hns", dest="nb_hns", type=int, default=15)
parser.add_argument("--pos-cutoff", dest="pos_cutoff", type=float, default=.75)
parser.add_argument("--neg-cutoff", dest="neg_cutoff", type=float, default=.35)
parser.add_argument("--val-size", dest="val_size", type=float, default=.1)
parser.add_argument("--bkg-dir", dest="bkg_dir", type=str, default="background")
parser.add_argument("--calc-pos-dir", dest="calc_pos_dir", type=str, default="calc_mal")
parser.add_argument("--calc-neg-dir", dest="calc_neg_dir", type=str, default="calc_ben")
parser.add_argument("--mass-pos-dir", dest="mass_pos_dir", type=str, default="mass_mal")
parser.add_argument("--mass-neg-dir", dest="mass_neg_dir", type=str, default="mass_ben")
parser.add_argument("--verbose", dest="verbose", action="store_true")
parser.add_argument("--no-verbose", dest="verbose", action="store_false")
parser.set_defaults(verbose=True)
args = parser.parse_args()
run_opts = dict(
target_height=args.target_height,
target_width=args.target_width,
segment_breast=args.segment_breast,
patch_size=args.patch_size,
nb_bkg=args.nb_bkg,
nb_abn=args.nb_abn,
nb_hns=args.nb_hns,
pos_cutoff=args.pos_cutoff,
neg_cutoff=args.neg_cutoff,
val_size=args.val_size,
bkg_dir=args.bkg_dir,
calc_pos_dir=args.calc_pos_dir,
calc_neg_dir=args.calc_neg_dir,
mass_pos_dir=args.mass_pos_dir,
mass_neg_dir=args.mass_neg_dir,
verbose=args.verbose
)
print "\n>>> Model training options: <<<\n", run_opts, "\n"
run(args.roi_mask_path_file, args.roi_mask_dir, args.pat_train_list_file,
args.full_img_dir, args.train_out_dir, args.val_out_dir, **run_opts)