-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataloader.py
312 lines (289 loc) · 13.6 KB
/
dataloader.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
# -*-coding:utf-8 -*-
'''
Created on 18/10/2022
@author: Carlos
'''
import numpy as np
import os
import nibabel as nib
from patcher import Patcher
import multiprocessing
def normalization(sigData, norm):
'''Normalize images with the method selected
Parameters:
- sigData: 4D numpy array of MRI signals or quantitative maps
- norm: normalization method
Return:
- Normalized image
- Normalization values
'''
np.seterr(invalid='ignore', divide='ignore')
maxims=[]
if norm=='max':
for vol in range(sigData.shape[-1]):
max1=np.max(sigData[:,:,:,vol])
maxims.append(max1)
sigData[:,:,:,vol]=sigData[:,:,:,vol]/max1
elif norm=='b0':
b0=sigData[:,:,:,0]
maxims.append(b0/b0)
for vol in range(1,sigData.shape[-1]):
maxims.append(b0)
sigData[:,:,:,vol]=sigData[:,:,:,vol]/b0
sigData=np.nan_to_num(sigData.astype('float32'), nan=0.0, posinf=0.0, neginf=0.0)
return sigData, maxims
def readADC(r,patch_s,norm,m,s,p2v):
'''Read dMRI simple quantitative parametric maps
Parameters
- r: folder path where the parametric maps are saved
- paths_s: pacth size
- norm: normalization method
- m: ROI binary mask file path
- s: stride for patcher
- p2v: 1 if patch2vox training
Return:
- Image as numpy array
'''
#Load qMRI parametric maps
lADCObj = nib.load(r+'/dri_highbval_low_ADC.nii')
lData=lADCObj.get_fdata()
lData=np.array(lData, 'float64')
hADCObj = nib.load(r+'/dri_lowbval_high_ADC.nii')
hData=hADCObj.get_fdata()
hData=np.array(hData, 'float64')
ls0Obj= nib.load(r+'/Festimate.nii')
#Calculate proxy F
fv=ls0Obj.get_fdata()
fv=np.array(fv, 'float64')
fv[np.isnan(fv)]=0
#Normalize the data
paramslist, maxims = normalization(np.stack((lData,hData,fv),axis=-1), norm)
#Load ROI mask
if m: mask=np.stack([np.array(nib.load(os.path.join(r.rsplit('/',1)[0],m)).get_fdata(), 'float64')]*paramslist.shape[-1],axis=-1)
else: mask=np.ones(paramslist.shape)
paramslist=paramslist*mask
#Patch the data if required
if patch_s: ADCjoin,divs,maxs,pad,m_p=Patcher(paramslist, np.stack([np.ones(hData.shape)]*paramslist.shape[-1],axis=-1), patch_size=patch_s,stride=s,p2v=p2v)
else: ADCjoin=np.reshape(paramslist,(-1,paramslist.shape[-1]))
return ADCjoin
def readParams(r,patch_s,norm, m,s,p2v):
'''Read DR-MRI advanced quantitative parametric maps
Parameters
- r: folder path where the parametric maps are saved
- paths_s: pacth size
- norm: normalization method
- m: ROI binary mask file path
- s: stride for patcher
- p2v: 1 if patch2vox training
Return:
- Image as numpy array
- NIfTI header
- Patches index
- Raw image shape
- Padding at each size
- Normalization values
- Quantitative parameters names
- ROI binary mask
'''
paramslist=[]
files=[]
#Load qMRI parametric maps
for file in os.listdir(r):
if 'dt' in file or 'dv' in file or 'kt' in file or 't2t' in file or 't2v' in file or 'fv' in file:
DtObj=nib.load(os.path.join(r,file))
DtData=np.array(DtObj.get_fdata(), 'float64')
paramslist.append(DtData)
files.append(file[-7:])
#Normalize the data
paramslist, maxims = normalization(np.stack(paramslist,axis=-1), norm)
#Load ROI mask
if m: mask=np.stack([np.array(nib.load(os.path.join(r.rsplit('/',1)[0],m)).get_fdata(), 'float64')]*paramslist.shape[-1],axis=-1)
else: mask=np.ones(paramslist.shape)
paramslist=paramslist*mask
#Patch the data if required
if patch_s: params,divs,maxs,pad,mask=Patcher(paramslist, mask, patch_size=patch_s,stride=s,p2v=p2v)
else:
maxs = DtData.shape
divs = pad = None
params=np.reshape(paramslist,(-1,paramslist.shape[-1]))
mask=np.reshape(mask, (-1,mask.shape[-1]))
return [params, DtObj, divs,maxs,pad,maxims, files, np.concatenate((mask,mask[:,:3]),axis=-1)]
def read21signals(inputlist):
'''Read MRI images of one patient to add to the signals intput and signals output of the models.
Parameters:
- path: DR-MRI file path
- patch_s: patch size
- norm: normalization method
- m: ROI binary mask file path
- s: stride for patcher
- p2v: 1 if patch2vox training
Return:
- Image as numpy array
- NIfTI header
- Patches index
- Raw image shape
- Padding at each size
- Normalization values
- ROI binary mask
'''
path, patch_s, norm, m,s, p2v = inputlist
del inputlist
#Load DR-MRI image
sigObj = nib.load(path)
sigData=sigObj.get_fdata()
sigData=np.array(sigData, 'float64')
#Load ROI mask
if m: mask=np.stack([np.array(nib.load(os.path.join(path.rsplit('/', 1)[0],m)).get_fdata(), 'float64')]*sigData.shape[-1],axis=-1)
else: mask=np.ones(sigData.shape)
sigData=sigData*mask
#Normalize the data
sigData, maxims = normalization(sigData, norm)
#Patch the data if required
if patch_s: res,divs,maxs,pad,mask=Patcher(sigData, mask, patch_size=patch_s,stride=s,p2v=p2v)
else:
maxs = sigData.shape
divs = pad = None
l=sigData.shape[0]*sigData.shape[1]*sigData.shape[2]
res=np.ones((l,1))
for dim4 in range(sigData.shape[3]):
layer=np.reshape(sigData[:,:,:,dim4], (l,1))
res=np.concatenate((res,layer),axis=1)
res=res[:,1:]
mask=np.reshape(mask, (-1,mask.shape[-1]))
return [res,sigObj,divs,maxs,pad, maxims,mask]
def dataloader_s2s(opt):
'''Load data for a signal-to-signal approach
Parameters:
-opt: dictionary of options from options.py
Return:
- Data sets: [[(x1,y1),(x2,y2),...,(xn,yn)]]
- ROI mask as numpy array
- Tuple: (patches index, raw input image shape, padding at each size, normalization values, NIfTI header, None)
'''
r=opt.input_rootname
r2=opt.input_rootname_test
if opt.mask: file='dri_ROI.nii'
else: file='dri_mppca5x5x1_den_unring_mocoaff_epicorr.nii'
print('')
print('Reading dataset...')
print('')
#Extraction of dataset (Features (X) and target (y) variables)
sets=[]
if not r.endswith('/'): r=r+'/'
if not r2.endswith('/'): r2=r2+'/'
leavedin = [c for c in os.listdir(r) if 'BL' in c and file in os.listdir(r+c) and all(t not in c for t in opt.test_case)]
leavedout = [c for c in os.listdir(r2) if 'BL' in c and any(t in c for t in opt.test_case) and file in os.listdir(r2+c)][0]
if opt.train:
print('Loading {} patients for training: {}\n'.format(len(leavedin),leavedin))
#Paralelize datareading
if not opt.patch2vox:
l=[[r+c+'/'+file, opt.patch_size, opt.norm, opt.mask, opt.stride, opt.patch2vox] for c in leavedin]
pool = multiprocessing.Pool(processes=opt.ncpus)
fparamsall = pool.map(read21signals, l)
pool.close()
fparams=fparamsall[0]; sig21=fparams[0]; mask_all = fparams[-1]
for p in fparamsall[1:]:
sig21 = np.concatenate((sig21,p[0]))
mask_all=np.concatenate((mask_all, p[-1]))
#Non-parallel datareading if algorithm pRFR (memory issues when parallel)
else:
for c in leavedin:
try:
fparams=read21signals([r+c+'/'+file, opt.patch_size, opt.norm, opt.mask, opt.stride, opt.patch2vox]) # Read the NIFTIs
if 'sig21' in locals(): sig21=np.concatenate((sig21,fparams[0])); mask_all=np.concatenate((mask_all, fparams[-1]))
else: sig21 = fparams[0]; mask_all = fparams[-1]
except Exception as e:
print('\nERROR while reading file from case {}'.format(c))
print(e)
print('Preprocessing data...\n')
if opt.patch_size and not opt.patch2vox: sig21 = np.stack([sig21[i,:,:,:,:] for i in range(sig21.shape[0]) if np.all(mask_all[i,:,:,:,:]!=0)],axis=0)
elif not opt.patch_size: sig21 = np.reshape(sig21[mask_all!=0],(-1,sig21.shape[-1])) #Delete zeros for training
np.random.shuffle(sig21) # Randomize voxels mantaining rows integrity
if opt.patch_size:
sig6=sig21[:,:,:,:,:opt.in_channels]
if opt.patch2vox: sig15=sig21[:,int(opt.patch_size/2),int(opt.patch_size/2),int(opt.patch_size/2),opt.in_channels:]
else: sig15=sig21[:,:,:,:,opt.in_channels:]
masked = []
else:
sig6=sig21[:,:opt.in_channels]
sig15=sig21[:,opt.in_channels:]
masked =[]
sets.append([(sig6[i], sig15[i]) for i in range(sig6.shape[0])])
else:
fparams=read21signals([r2+leavedout+'/'+file, opt.patch_size, opt.norm, opt.mask, opt.stride, opt.patch2vox])
if opt.patch_size:
caseout6s=fparams[0][:,:,:,:,:opt.in_channels] # Read test input
if opt.patch2vox: caseout15s=fparams[0][:,int(opt.patch_size/2),int(opt.patch_size/2),int(opt.patch_size/2),opt.in_channels:] # Read test output
else: caseout15s=fparams[0][:,:,:,:,opt.in_channels:] # Read test output
masked = np.array(read21signals([r2+leavedout+'/'+file, 0, 'non', opt.mask, opt.stride, opt.patch2vox])[0]!=0, 'float64')
else:
params=np.reshape(fparams[0][fparams[-1]!=0],(-1,fparams[0].shape[-1]))
caseout6s=params[:,:opt.in_channels] # Read test input
caseout15s=params[:,opt.in_channels:] # Read test output
masked = np.array(fparams[-1][:,opt.in_channels:]!=0, 'float64') # Save mask as array
sets.append([(caseout6s[i], caseout15s[i]) for i in range(caseout6s.shape[0])])
return sets, masked, (fparams[2], fparams[3], fparams[4], fparams[5], fparams[1], None)
def dataloader_m2m(opt):
'''Load data for a maps-to-maps approach
Parameters:
- opt: dictionary of options from options.py
Return:
- Data sets: [[(x1,y1),(x2,y2),...,(xn,yn)]]
- ROI mask as numpy array
- Tuple: (patches index, single image input shape, padding at each size, normalization values, NIfTI header, ROI mask)
'''
r=opt.input_rootname
r2=opt.input_rootname_test
if opt.mask: folders=['dri_ADC_ROI', 'dri_newregnosigmat2ivimkurt_ROI']
else: folders=['dri_ADC', 'dri_newregnosigmat2ivimkurt']
print('')
print('Reading dataset...')
print('')
#Extraction of dataset (Features (X) and answers (y) variables)
sets=[]
if not r.endswith('/'): r=r+'/'
if not r2.endswith('/'): r2=r2+'/'
leavedin = [c for c in os.listdir(r) if 'BL' in c and folders[0] in os.listdir(r+c) and folders[1] in os.listdir(r+c) and all(t not in c for t in opt.test_case)]
leavedout = [c for c in os.listdir(r2) if 'BL' in c and any(t in c for t in opt.test_case) and folders[0] in os.listdir(r2+c) and folders[1] in os.listdir(r2+c)][0]
if opt.train:
print('Loading {} patients for training: {}\n'.format(len(leavedin),leavedin))
for c in leavedin:
try:
read=readADC(r+c+'/'+folders[0], opt.patch_size,opt.norm, opt.mask, s=opt.stride, p2v=opt.patch2vox)
if 'ADC' in locals(): ADC=np.concatenate((ADC, read))
else: ADC = read
fparams=readParams(r+c+'/'+folders[1], opt.patch_size, opt.norm, opt.mask, opt.stride, opt.patch2vox)
if 'params' in locals(): params=np.concatenate((params,fparams[0])); mask_all=np.concatenate((mask_all, fparams[-1]))
else: params = fparams[0]; mask_all = fparams[-1]
except Exception as e:
print('')
print('ERROR while reading file from case {}'.format(c))
print(e)
ADCparams=np.concatenate((ADC,params),axis=-1) #Join inputs and output in one matrix
if opt.patch_size: ADCparams = np.stack([ADCparams[i,:,:,:,:] for i in range(ADCparams.shape[0]) if np.all(mask_all[i,:,:,:,:]!=0)],axis=0) #Delete patches that are not full inside ROI
else: ADCparams=np.reshape(ADCparams[mask_all!=0],(-1,ADCparams.shape[-1])) #Delete zeros for training
np.random.shuffle(ADCparams) #Shuffle mantaining rows integrity
if opt.patch_size:
ADC=ADCparams[:,:,:,:,:opt.in_channels]
if opt.patch2vox: params=ADCparams[:,int(opt.patch_size/2),int(opt.patch_size/2),int(opt.patch_size/2),opt.in_channels:]
else: params=ADCparams[:,:,:,:,opt.in_channels:]
masked = []
else:
ADC=ADCparams[:,:opt.in_channels]
params=ADCparams[:,opt.in_channels:]
masked =[]
sets.append([(ADC[i], params[i]) for i in range(ADC.shape[0])])
else:
ADC=readADC(r2+leavedout+'/'+folders[0], opt.patch_size,opt.norm, opt.mask, s=opt.stride, p2v=opt.patch2vox) # Read test input
fparams=readParams(r2+leavedout+'/'+folders[1], opt.patch_size,opt.norm,opt.mask, opt.stride, opt.patch2vox) # Read test output
if opt.patch_size:
masked = np.array(readParams(r2+leavedout+'/'+folders[1], 0,'non',opt.mask, opt.stride, opt.patch2vox)[0]!=0, 'float64')
ADCf=ADC
if opt.patch2vox: params=fparams[0][:,int(opt.patch_size/2),int(opt.patch_size/2),int(opt.patch_size/2),:] # Read test output
else: params=fparams[0]
else:
ADCf=np.reshape(ADC[fparams[-1][:,:opt.in_channels]!=0],(-1,opt.in_channels))
params=np.reshape(fparams[0][fparams[-1][:,opt.in_channels:]!=0],(-1,opt.out_channels))
masked = np.array(fparams[-1][:,opt.in_channels:]!=0, 'float64') # Save mask as array
sets.append([(ADCf[i], params[i]) for i in range(ADCf.shape[0])])
return sets, masked, (fparams[2], fparams[3], fparams[4], fparams[5], fparams[1],fparams[6])