/
detectSeg3DkerasDR_gh.py
451 lines (345 loc) · 20.8 KB
/
detectSeg3DkerasDR_gh.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
def detectSeg3DkerasDR(trainMode,testMode,params):
print('21!')
import os, sys
sys.path.insert(0, '/add-directory-path-where-needed/this-folder')
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
import funcs_gh
import matplotlib.pyplot as plt
import pickle
from datetime import datetime, timedelta
import numpy as np
from keras.models import Model,load_model,Sequential
from keras.layers import Reshape,Input, concatenate, Conv3D,Dense,TimeDistributed
from keras.layers import MaxPooling3D, UpSampling3D, LSTM,ConvLSTM2D
from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint,TensorBoard
from keras import backend as K
from keras.preprocessing.image import ImageDataGenerator
from matplotlib import pyplot
import tensorflow as tf
from networks_gh import get_unet2, get_unet3, get_rbunet, get_meshNet,get_denseNet,calculatedPerfMeasures
from networks_gh import get_unetCnnRnn, augment_sample
from networks_gh import get_denseNet103, augmentation, generateAugmentation, IoU3D
from selectTrainAndTestSubjects_gh import selectTrainAndTestSubjects
from scipy.ndimage import zoom
from scipy import signal
from skimage import morphology
from skimage import data
from skimage.feature import corner_harris, corner_subpix, corner_peaks
from skimage.transform import warp, AffineTransform
import pandas as pd
TestSetNum=params['TestSetNum'];
fileNumModel=params['fileNumModel'];
tDim=params['tDim'];
tpUsed=params['tpUsed'];
PcUsed=params['PcUsed'];
deepRed=params['deepReduction'];
#multiHead=params['multiHead'];
if PcUsed:
tDim=5;
xDim=64; yDim=64; zDim=64;
xyDim = 64;
dsFactor = 3.5;
n_channels = tDim;
n_classes = 3 # (background, left-kidney, right-kidney)
if PcUsed==1:
pc='pc/';
elif PcUsed==2:
pc='kpc/';
elif PcUsed==3:
pc='tsne/';
elif PcUsed==4:
pc='/'
net=params['networkToUse'];
if net=='meshNet':
xyDim=96;zDim=96;
############ stratify between train and test data
subjectNamesNormalTrain, subjectNamesNormalTest, _ ,testKidCond ,subjectBaselinesTest = selectTrainAndTestSubjects(TestSetNum);
############ generate train batch data
def generate_batch():
for samples in generate_samples():
label_batch=np.zeros((len(samples),zDim,xyDim,xyDim,3))
boxedLabel_batch=np.zeros((len(samples),zDim,xyDim,xyDim,3))
image_batch=np.zeros((len(samples),zDim,xyDim,xyDim,tDim))
for s in range(len(samples)):
data4D = pickle.load(open("/path-to-folder-containing-downsampled-images-for-detection-model/singleSubjectsV4pc_detect/"+subjectNamesNormalTrain[samples[s]]+".p","rb" ));
da=data4D[subjectNamesNormalTrain[samples[s]]+'D'].T;
im0 = zoom(da,(1,zDim/da.shape[1],1/dsFactor,1/dsFactor),order=0);
la0 = zoom(data4D[subjectNamesNormalTrain[samples[s]]+'M'].T,(zDim/da.shape[1],1/dsFactor,1/dsFactor),order=0);
Labels=la0[:,:,:,np.newaxis].astype(int);
Labels[Labels>2]=2;
boxedlab = np.copy(Labels)
uniq=np.unique(Labels);
for i in range(1,len(uniq)):
lNdx=np.where(Labels==uniq[i]);
boxedlab[min(lNdx[0]):max(lNdx[0]),min(lNdx[1]):max(lNdx[1]),min(lNdx[2]):max(lNdx[2])]=uniq[i];
lRb=np.copy(boxedlab);lLb=np.copy(boxedlab);lBb=np.copy(boxedlab);
lRb[lRb!=1]=0;lLb[lLb!=2]=0;lLb[lLb!=0]=1;lBb[lBb==0]=5;lBb[lBb!=5]=0;lBb[lBb==5]=1;
labelsBoxed=np.concatenate((lBb,lRb,lLb),axis=3);
lR=np.copy(Labels);lL=np.copy(Labels);lB=np.copy(Labels);
lR[lR!=1]=0;lL[lL!=2]=0;lL[lL!=0]=1;lB[lB==0]=5;lB[lB!=5]=0;lB[lB==5]=1;
labels=np.concatenate((lB,lR,lL),axis=3);
im1=np.swapaxes(im0.T,0,2);
image_batch[s,:,:,:,:]= im1;
label_batch[s,:,:,:,:] = labels;
boxedLabel_batch[s,:,:,:,:] = labelsBoxed;
# add data augmentation
for i in range(image_batch.shape[0]):
image_batch[i], label_batch[i],boxedLabel_batch[i] = augment_sample(image_batch[i], label_batch[i],boxedLabel_batch[i])
yield(image_batch, label_batch,boxedLabel_batch)
n_samples = 46 # number of train subjects
batch_size = 2 #3 #4 #batch size
n_batches = int(n_samples/batch_size); # number of batches
def generate_samples():
sample_ids = np.random.permutation(n_samples)
for i in range(n_batches):
inds = slice(i*batch_size, (i+1)*batch_size)
yield sample_ids[inds]
################### visual check for cropped segmentaation
#for i in range(40,42):
# # plt.figure();plt.imshow(Data[i,:,:,10,0].T);
# f, axarr = plt.subplots(1, 2);
#
# # axarr[0].imshow(DataTest[0,zFix,:,:,1].T);
# axarr[0].imshow(Data[i,:,:,40,2].T);
# axarr[1].imshow(Labels[i,:,:,40,0].T);
########################## set weights,train and test ##############################
### dice accuracy
def dice_coef(y_true, y_pred):
y_true_f = y_true.flatten();
y_pred_f = y_pred.flatten()
intersection = np.sum(y_true_f * y_pred_f)
return (2. * intersection) / (np.sum(y_true_f) + np.sum(y_pred_f))
### tversky_coefficient
def tversky_coef(y_true, y_pred, alpha, beta, smooth=1):
y_true_f = K.flatten(y_true)
y_true_f_r = K.flatten(1. - y_true)
y_pred_f = K.flatten(y_pred)
y_pred_f_r = K.flatten(1. - y_pred)
weights = 1.
intersection = K.sum(y_pred_f * y_true_f * weights)
fp = K.sum(y_pred_f * y_true_f_r)
fn = K.sum(y_pred_f_r * y_true_f * weights)
return (intersection + smooth) / (intersection + alpha * fp + beta * fn + smooth)
### tversky_coef loss
def tversky_loss(alpha, beta, weights=False):
def tversky(y_true, y_pred):
return -tversky_coef(y_true, y_pred, alpha, beta, weights)
return tversky
### tversky_coef loss
tversky = tversky_loss(alpha=0.3, beta=0.7, weights=False)
# initial class weights
class_weights=np.array([0.05,0.6,0.6]);
if net=='rbUnet':
model = get_rbunet(xyDim,zDim,n_channels,n_classes,deepRed,0);
elif net=='Unet':
model = get_unet3(xyDim,zDim,n_channels,n_classes,deepRed,0);
elif net=='Unet-rnn':
predLayersCat=get_unetCnnRnn(xyDim,zDim,1,n_classes,deepRed,0);
predLayersCat=Reshape((1,64, 64, 64))(predLayersCat)
for i in range(1,n_channels):
predLayer = get_unetCnnRnn(xyDim,zDim,1,n_classes,deepRed,0);
predLayer=Reshape((1,64, 64, 64))(predLayer)
predLayersCat= concatenate([predLayersCat, predLayer], axis=1)
model = Sequential()
model.add(ConvLSTM2D(64, kernel_size=3, padding='same',input_shape=(50, 64, 64,64)))
#lstm_num_predictions=1;
#model.add(Dense(lstm_num_predictions))
model = Model(inputs=[inputs], outputs=[Pred])
model.compile(optimizer=Adam(lr=1e-4), loss=dice_coef_loss2)
elif net=='meshNet':
xyDim=96;zDim=96;
model = get_meshNet(xyDim,zDim,n_channels,n_classes,deepRed,0);
elif net== 'denseNet':
model = get_denseNet(xyDim,zDim,n_channels,n_classes,deepRed,0);
elif net== 'tNet':
model = get_denseNet103(xyDim,zDim,n_channels,n_classes,deepRed,0);
### create folder to hold detection model(s)
if isinstance(fileNumModel, (int)):
fileNumModel=str(fileNumModel);
else:
fileNumModel='Net'+net+'_time'+str(tDim)+'_pcUsed'+str(PcUsed)+'_tpUsed'+str(tpUsed)+'_DR'+str(deepRed)+'_testSet'+str(TestSetNum);
address = "path-to-folder-to-hold-detection-model(s)/"+fileNumModel+"/"
if trainMode:
os.system('mkdir '+address);
#current_time = datetime.now() + timedelta(hours=-5)
#log_dir=address+str(current_time)[:19]
#callbacks = [
#TensorBoard(address+'tbevents',histogram_freq=0, write_graph=True, write_images=False),
#ModelCheckpoint(address+'conv3Dkeras.h5',verbose=1,monitor='val_loss', save_best_only=True, save_weights_only=True),
#]
# set multiHead
multiHead = 0;
nb_epoch = 400; #500 #100
epCounter = 0;
for e in range(nb_epoch):
print("epoch %d" % e)
for image_batch, label_batch,boxedLbatch in generate_batch():
print(epCounter,label_batch.shape[0]);
#epCounter+=(batch_size+1);
xx=(batch_size*2)+1;
epCounter+=xx;
if multiHead == 1:
labelsDict={'seg':label_batch,'box':boxedLbatch};
classWdict={'seg':class_weights,'box':class_weights};
else:
labelsDict={'seg':label_batch};
classWdict={'seg':class_weights};
model.fit(image_batch, label_batch, batch_size=batch_size*2, class_weight=class_weights,
initial_epoch =epCounter, epochs=epCounter+xx,verbose=1, shuffle=True,validation_split=0.5); #,callbacks=callbacks);
if epCounter >= 100 and epCounter <= 500:
model.save(address+'detect3D_'+str(epCounter)+'.h5')
if epCounter >= 2000 and epCounter <= 3000:
model.save(address+'detect3D_'+str(epCounter)+'.h5')
if epCounter >= 5000 and epCounter <= 10000:
model.save(address+'detect3D_'+str(epCounter)+'.h5')
if epCounter >= 15000 and epCounter <= 20000:
model.save(address+'detect3D_'+str(epCounter)+'.h5')
if epCounter >= 28000 and epCounter <= 35000:
model.save(address+'detect3D_'+str(epCounter)+'.h5')
if epCounter >= 40000:
model.save(address+'detect3D_'+str(epCounter)+'.h5')
performanceMeasures,avgPerf,missedVoxels=[],[],[];
if testMode == 1:
fileNumModel='Net'+net+'_time'+str(tDim)+'_pcUsed'+str(PcUsed)+'_tpUsed'+str(tpUsed)+'_DR'+str(deepRed)+'_testSet'+str(TestSetNum);
address = "path-to-folder-containing-trained-detection-model(s)/"+fileNumModel+"/"
#address = "path-to-folder-to-hold-detection-model(s)" + "/NetrbUnet_time5_pcUsed1_tpUsed50_DR0_testSet1/"
#address = "path-to-folder-to-hold-detection-model(s)" + "/NetrbUnet_time5_pcUsed1_tpUsed50_DR0_testSet2/"
selectedEpoch=params['selectedEpoch'];
# if isinstance(selectedEpoch, (int)):
# selectedEpoch=str(selectedEpoch);
# else:
# txt_file = open(address+'selectedEpoc.txt','r')
# selectedEpoch=str(int(txt_file.read()))
model.load_weights(address+'detect3D_'+selectedEpoch+'.h5');
# perform detection (coarse segmentation) for each test subject
# and save to a file
for s in range(len(subjectNamesNormalTest)):
xyDim = 64; zDim = 64;
LabelsTest=np.zeros((1,zDim,xyDim,xyDim));
DataTest=np.zeros((1,zDim,xyDim,xyDim,tDim));
boxTest=np.zeros((1,12));
# access test input data
data4D = pickle.load(open("/path-to-folder-containing-downsampled-images-for-detection-model/"+"singleSubjectsV4pc_detect/"+subjectNamesNormalTest[s]+".p","rb" ));
#### extract ground-truth labels and 'box' to evaluate performance of detection model
da = data4D[subjectNamesNormalTest[s]+'D'].T;
im0 = zoom(da,(1,zDim/da.shape[1],1/dsFactor,1/dsFactor),order=0);
la0 = zoom(data4D[subjectNamesNormalTest[s]+'M'].T,(zDim/da.shape[1],1/dsFactor,1/dsFactor),order=0);
B=data4D[subjectNamesNormalTest[s]+'B'];
B[:,[0,1,3,4]]=(B[:,[0,1,3,4]]/dsFactor).astype(int);
B[:,[2,5]]=(B[:,[2,5]]*(zDim/da.shape[1])).astype(int);
sx = 0;
LabelsTest[sx,0:np.size(la0,0),:,:]=la0;
DataTest[sx,:,:,:,:]=np.swapaxes(im0.T,0,2);
boxTest[sx,:]=B.flatten();
#### perform prediction ####
imgs_mask_test= model.predict(DataTest, verbose=1);
multiHead = 0;
if multiHead:
labels_pred=np.argmax(imgs_mask_test[0], axis=4)
else:
labels_pred=np.argmax(imgs_mask_test, axis=4)
# ensure all detected labels for right kidney are on the right half of x dimension
labels_pred[:,:,:,0:int(xyDim/2)][labels_pred[:,:,:,0:int(xyDim/2)]==2]=1;
labels_pred[:,:,:,int(xyDim/2):][labels_pred[:,:,:,int(xyDim/2):]==1]=2;
#LabelsTest2=np.copy(LabelsTest);labels_pred2=np.copy(labels_pred);
#LabelsTest2[LabelsTest2!=0]=1;labels_pred2[labels_pred2!=0]=1;
##### generate bounding boxes from coarse segmentation #####
boxPred=np.zeros((DataTest.shape[0],12))
si = 0;
aL=np.nonzero(labels_pred[si,:,:,:].T==2); #left
aR=np.nonzero(labels_pred[si,:,:,:].T==1); #right
left = labels_pred[si,:,:,:].T==2;
left = left.astype(int);
right = labels_pred[si,:,:,:].T==1;
right = right.astype(int);
if aL[0].size!=0:
boxL=np.array([int((min(aL[0])+max(aL[0]))/2),int((min(aL[1])+max(aL[1]))/2),int((min(aL[2])+max(aL[2]))/2),\
(max(aL[0])-min(aL[0])),(max(aL[1])-min(aL[1])),(max(aL[2])-min(aL[2]))])
else:
boxL=np.zeros((6,));
if aR[0].size!=0:
boxR=np.array([int((min(aR[0])+max(aR[0]))/2),int((min(aR[1])+max(aR[1]))/2),int((min(aR[2])+max(aR[2]))/2),\
(max(aR[0])-min(aR[0])),(max(aR[1])-min(aR[1])),(max(aR[2])-min(aR[2]))])
else:
boxR=np.zeros((6,));
boxes=np.hstack([np.array(boxR),np.array(boxL)]);
boxPred[si,:]=boxes;
# #### calculate number of missed voxels (number of voxels out of bounding box)
# columns = ['Name','kidney Condition','F1-Score', 'Prec','Rec','VEE','testSet','Model','IoU','missedVoxelsR','missedVoxelsL'];
# index=np.arange(1);
# performanceMeasures= pd.DataFrame(index=index, columns=columns)
# performanceMeasures= performanceMeasures.fillna(0);
# IoU,missedVoxels=IoU3D(boxPred[si,:],boxTest[si,:],labels_pred2[si,:,:,:].T);
# avgPerfOverKidneys=calculatedPerfMeasures(LabelsTest2[si,:,:,:],labels_pred2[si,:,:,:]);
# performanceMeasures.ix[s]=pd.Series({'Name':subjectNamesNormalTest[s],'kidney Condition':testKidCond[s],'F1-Score':avgPerfOverKidneys[0]*100,'Prec':avgPerfOverKidneys[1]*100,\
# 'Rec':avgPerfOverKidneys[2]*100,'VEE':avgPerfOverKidneys[4],'testSet':TestSetNum,'Model':net,'IoU':IoU,'missedVoxelsR':missedVoxels[0],'missedVoxelsL':missedVoxels[1]});
##### save detection results as coarse segmentation masks for each test subject ####
pathToFolder = 'path-to-folder-to-contain-segmented-image-files' + '/detected/' + 'subjectNamesNormalTest[s]' + '_seq1';
if not os.path.exists(pathToFolder):
os.makedirs(pathToFolder)
# extract test image volume
# fileAddress = path to .xls sheet that contains time information for each test subject file
# fileAddress='path-to-folder"+"/subjectDicomInfo_gh.xls';
# subjectInfo=pd.read_excel(fileAddress, sheetname=0);
# reconMethod='SCAN';
# vol4D00,_,_,_,_ = funcs_gh.readData4(subjectNamesNormalTest[s],subjectInfo,reconMethod,0);
# extract test image volume dimensions
vol4D00 = data4D[subjectNamesNormalTest[s]+'D'];
zDimm = vol4D00.shape[2];
xyDim = 224;
# resample predicted labels (right, left) to original test image spatial dimensions
KMR = zoom(right,(xyDim/np.size(right,0),xyDim/np.size(right,1),zDimm/np.size(right,2)),order=0);
KML = zoom(left,(xyDim/np.size(left,0),xyDim/np.size(left,1),zDimm/np.size(left,2)),order=0);
predMaskR=np.zeros((1,xyDim,xyDim,zDimm));
predMaskL=np.zeros((1,xyDim,xyDim,zDimm));
sc = 0;
predMaskR[sc,:,:,:] = KMR;
predMaskL[sc,:,:,:] = KML;
if np.sum(predMaskR) != 0:
predMaskL=morphology.remove_small_objects(predMaskL.astype(bool), min_size=256,in_place=True).astype(int);
if np.sum(predMaskL) != 0:
predMaskR=morphology.remove_small_objects(predMaskR.astype(bool), min_size=256,in_place=True).astype(int);
Masks2Save={};
predMaskR2=zoom(predMaskR[sc,:,:,:],(1,1,1),order=0);
predMaskL2=zoom(predMaskL[sc,:,:,:],(1,1,1),order=0);
Masks2Save['R']=np.copy(predMaskR2.astype(float));
Masks2Save['L']=np.copy(predMaskL2.astype(float));
# path to .xls sheet that contains time information for each test subject file (subjectNamesNormalTest[s])
fileAddress='path-to-folder"+"/subjectDicomInfo_gh.xls';
subjectInfo=pd.read_excel(fileAddress, sheetname=0);
reconMethod='SCAN';
# write coarse segmentation (detection) results to file
funcs_gh.writeMasksDetect(subjectNamesNormalTest[s],subjectInfo,reconMethod,Masks2Save,1);
"""
#### calculate number of missed voxels (number of voxels out of bounding box)
columns = ['Name','kidney Condition','F1-Score', 'Prec','Rec','VEE','testSet','Model','IoU','missedVoxelsR','missedVoxelsL'];
index=np.arange(len(subjectNamesNormalTest));
performanceMeasures= pd.DataFrame(index=index, columns=columns)
performanceMeasures= performanceMeasures.fillna(0);
for s in range(DataTest.shape[0]):
IoU,missedVoxels=IoU3D(boxPred[s,:],boxTest[s,:],labels_pred2[s,:,:,:].T);
avgPerfOverKidneys=calculatedPerfMeasures(LabelsTest2[s,:,:,:],labels_pred2[s,:,:,:]);
performanceMeasures.ix[s]=pd.Series({'Name':subjectNamesNormalTest[s],'kidney Condition':testKidCond[s],'F1-Score':avgPerfOverKidneys[0]*100,'Prec':avgPerfOverKidneys[1]*100,\
'Rec':avgPerfOverKidneys[2]*100,'VEE':avgPerfOverKidneys[4],'testSet':TestSetNum,'Model':net,'IoU':IoU,'missedVoxelsR':missedVoxels[0],'missedVoxelsL':missedVoxels[1]});
normalPerf=performanceMeasures[performanceMeasures['kidney Condition'] == 'N'].iloc[:,2:6].mean().tolist();
abnormalPerf=performanceMeasures[performanceMeasures['kidney Condition'] == 'A'].iloc[:,2:6].mean().tolist();
avgPerf=normalPerf+abnormalPerf
if 1:
volumEstimError=np.zeros((labels_pred.shape[0],))
performanceMeasures=np.zeros((len(subjectNamesNormalTest),));
for s in range(DataTest.shape[0]):
zAx=40;
f, axarr = plt.subplots(2, 3);
#axarr[0].imshow(imgs_mask_test[0,:,:,0].T,cmap='gray');
axarr[0,0].imshow(labels_pred[s,zAx,:,:],cmap='gray');axarr[0,0].set_title('Predicted');
axarr[0,1].imshow(LabelsTest[s,zAx,:,:],cmap='gray');axarr[0,1].set_title('Original');
axarr[0,2].imshow(DataTest[s,zAx,:,:,1],cmap='gray');axarr[0,2].set_title(subjectNamesNormalTest[s]);
axarr[1,0].imshow(labels_pred[s,:,int(xyDim/2),:],cmap='gray');
axarr[1,1].imshow(LabelsTest[s,:,int(xyDim/2),:],cmap='gray');
axarr[1,2].imshow(DataTest[s,:,int(xyDim/2),:,1],cmap='gray');
volumEstimError[s]=np.count_nonzero(LabelsTest2[s,:,:,:])-np.count_nonzero(labels_pred2[s,:,:,:]);
performanceMeasures[s]=dice_coef(LabelsTest2[s,:,:,:],labels_pred2[s,:,:,:]);
"""
return performanceMeasures,avgPerf