def get_val_dice(cnn, config):
val_dsc = []
taskdir = cfg['taskdir']
# load array from selected val to test dsc and compare with cart vw label
if os.path.exists(taskdir + '/valarray_cart.npy'):
xarray_val = np.load(taskdir + '/valarray_cart.npy')
xlabel_val = np.load(taskdir + '/vallabel_cart.npy')
else:
val_exams = dbloader.list_exams('val', shuffle=False)
xarray_val = []
xlabel_val = []
bi = 0
for ei in val_exams:
caseloader = CaseLoader(ei)
print(caseloader)
for slicei in caseloader.slices:
if caseloader.valid_slicei(slicei) == False:
continue
if bi % 20 == 0:
polarstack = caseloader.loadstack(slicei,
'cart_patch',
nei=config['depth'] // 2)
xarray_val.append(polarstack[..., None])
cartabel = caseloader.loadstack(slicei,
'cart_label',
nei=config['depth'] //
2)[...,
config['depth'] // 2]
xlabel_val.append(cartabel[:, :, 1] - cartabel[:, :, 0])
bi += 1
xarray_val = np.array(xarray_val)
xlabel_val = np.array(xlabel_val)
print(xarray_val.shape, xlabel_val.shape)
np.save(taskdir + '/valarray_cart.npy', xarray_val)
np.save(taskdir + '/vallabel_cart.npy', xlabel_val)
polarbd_all = cnn.predict(xarray_val,
batch_size=config['G'] * config['batchsize'])
for ti in range(xarray_val.shape[0]):
carlabel = xlabel_val[ti]
polarimg = xarray_val[ti]
polarbd = polarbd_all[ti] * config['width'] // 2
# polarprob = polarprobr[:,:,0,1]-polarprobr[:,:,0,0]
contourin, contourout = toctbd(polarbd, config['patchheight'],
config['patchheight'])
carseg = plotct(polarimg.shape[1], contourin, contourout)
cdsc = DSC(carlabel, carseg)
val_dsc.append(cdsc)
# pxconf = (np.sum(polarprob[polarseg>0])-np.sum(polarprob[polarseg==0]))/np.sum(polarseg>0)
# sconf.append(pxconf)
print('\rPred',
ti,
'/',
xarray_val.shape[0],
'val dice',
'%.5f' % np.mean(val_dsc),
end="")
return np.mean(val_dsc)
def data_generator(config, exams, aug):
xarray = np.zeros([
config['batchsize'], config['height'], config['width'],
config['depth'], config['channel']
])
yarray = np.zeros([config['batchsize'], config['patchheight'], 2])
bi = 0
while 1:
for ei in exams:
caseloader = CaseLoader(ei)
# print(caseloader)
for slicei in caseloader.slices:
polarstack = caseloader.loadstack(slicei,
'cart_patch',
nei=config['depth'] // 2)
if polarstack is None:
# print('skip',caseloader,slicei)
continue
polarcont = caseloader.loadstack(slicei,
'polar_cont',
nei=config['depth'] // 2)
polar_cont_center = polarcont[:, :, config['depth'] // 2]
xarray[bi] = polarstack[..., None]
yarray[bi] = polar_cont_center
bi += 1
if bi == config['batchsize']:
bi = 0
if aug == True:
for offi in random.sample(range(config['patchheight']),
config['rottimes']):
xarray_off = batch_polar_rot(xarray, offi)
yarray_off = batch_polar_rot(yarray, offi)
yield (xarray_off, yarray_off)
else:
yield xarray, yarray
def data_generator(config, exams, aug):
xarray = np.zeros([
config['batchsize'], config['height'], config['width'],
config['depth'], config['channel']
])
yarray = np.zeros([config['batchsize'], config['patchheight'], 2])
bi = 0
while 1:
for ei in exams:
caseloader = CaseLoader(ei)
# print(caseloader)
for slicei in caseloader.slices:
polarstack = caseloader.loadstack(slicei,
'cart_patch',
nei=config['depth'] // 2)
if polarstack is None:
# print('skip',caseloader,slicei)
continue
polarstack = cv2.resize(polarstack, (256, 256))
polarcont = caseloader.loadstack(slicei,
'polar_cont',
nei=config['depth'] // 2)
polar_cont_center = polarcont[:, :, config['depth'] // 2]
xarray[bi] = polarstack[..., None]
yarray[bi] = polar_cont_center
bi += 1
if bi == config['batchsize']:
bi = 0
if aug == True:
for offi in range(config['rottimes']):
rotsarr = []
for i in range(xarray.shape[0]):
rots = np.arange(4)
random.shuffle(rots)
rotsarr.append(rots)
rotsarr = np.array(rotsarr)
for roti in range(4):
xarray_off, yarray_off = batch_cart_rot(
xarray, yarray, rotsarr[:, roti])
yield (xarray_off, yarray_off)
else:
yield (xarray, yarray)
def data_generator(config, exams, aug):
xarray = np.zeros([
config['batchsize'], config['height'], config['width'],
config['depth'], config['channel']
])
yarray = np.zeros([config['batchsize'], config['patchheight'], 2])
bi = 0
while 1:
for ei in exams:
caseloader = CaseLoader(ei)
#print(caseloader)
for slicei in caseloader.slices:
if caseloader.valid_slicei(slicei) == False:
continue
if caseloader.caselist['slices'][slicei]['augpatch'] == []:
continue
gen_aug_patch(caseloader, slicei)
print(caseloader, slicei, 'gen')
aug_polar_patch_batch = caseloader.load_aug_patch(
slicei, 'polar_patch')
aug_polar_cont_batch = caseloader.load_aug_patch(
slicei, 'polar_cont')
for augi in range(len(aug_polar_patch_batch)):
aug_cart_patch = aug_polar_patch_batch[augi]
xarray[bi] = aug_polar_patch_batch[augi, :, :, :, None]
aug_polar_cont = aug_polar_cont_batch[augi]
yarray[bi] = aug_polar_cont_batch[augi]
bi += 1
if bi == config['batchsize']:
bi = 0
if aug == True:
for offi in random.sample(range(config['height']),
config['rottimes']):
xarray_off = batch_polar_rot(xarray, offi)
yarray_off = batch_polar_rot(yarray, offi)
yield (xarray_off, yarray_off)
else:
yield (xarray, yarray)
def data_generator(config, exams, aug):
xarray = np.zeros([config['batchsize'], config['height'], config['width'], config['depth'], config['channel']])
yarray = np.zeros([config['batchsize'], config['patchheight'], 2])
bi = 0
while 1:
for ei in exams:
caseloader = CaseLoader(ei)
# print(caseloader)
for slicei in caseloader.slices:
if caseloader.valid_slicei(slicei) == False:
continue
if caseloader.caselist['slices'][slicei]['augpatch'] == [] :
continue
gen_aug_patch(caseloader, slicei)
print(caseloader, slicei, 'gen')
aug_cart_patch_batch = caseloader.load_aug_patch(slicei, 'cart_patch')
aug_polar_cont_batch = caseloader.load_aug_patch(slicei, 'polar_cont')
for augi in range(len(aug_cart_patch_batch)):
aug_cart_patch = cv2.resize(aug_cart_patch_batch[augi], (config['width'], config['height']))
xarray[bi] = aug_cart_patch[..., None]
aug_polar_cont = aug_polar_cont_batch[augi]
yarray[bi] = aug_polar_cont_batch[augi]
bi += 1
if bi == config['batchsize']:
bi = 0
if aug == True:
for offi in range(config['rottimes']):
rotsarr = []
for i in range(xarray.shape[0]):
rots = np.arange(4)
random.shuffle(rots)
rotsarr.append(rots)
rotsarr = np.array(rotsarr)
for roti in range(4):
xarray_off, yarray_off = batch_cart_rot(xarray, yarray, rotsarr[:, roti])
yield (xarray_off, yarray_off)
else:
yield (xarray, yarray)
def data_generator(config, exams, aug):
xarray = np.zeros([
config['batchsize'], config['height'], config['width'],
config['depth'], config['channel']
])
yarray = np.zeros(
[config['batchsize'], config['height'], config['width'], 1, 1])
bi = 0
while 1:
for ei in exams:
caseloader = CaseLoader(ei)
# print(caseloader)
for slicei in caseloader.slices:
if caseloader.valid_slicei(slicei) == False:
continue
#load center patch
xarray[bi] = caseloader.loadstack(slicei,
'cart_patch')[:, :, :, None]
yarray[bi] = caseloader.load_cart_vw(slicei)[:, :, None, None]
bi += 1
if bi == config['batchsize']:
bi = 0
yield (xarray, yarray)
# load aug patch
aug_cart_patch_batch = caseloader.load_aug_patch(
slicei, 'cart_patch')
aug_cart_label_batch = caseloader.load_aug_patch(
slicei, 'cart_label')
for augi in range(len(aug_cart_patch_batch)):
xarray[bi] = aug_cart_patch_batch[augi, :, :, :, None]
yarray[bi] = aug_cart_label_batch[augi][:, :, None, None]
bi += 1
if bi == config['batchsize']:
bi = 0
yield (xarray, yarray)
def data_generator(config, exams, aug):
xarray = np.zeros([
config['batchsize'], config['height'], config['width'],
config['depth'], config['channel']
])
yarray = np.zeros([config['batchsize'], config['patchheight'], 2])
bi = 0
while 1:
for ei in exams:
caseloader = CaseLoader(ei)
# print(caseloader)
for slicei in caseloader.slices:
if caseloader.valid_slicei(slicei) == False:
continue
cartcont = caseloader.load_cart_cont(slicei)
min_dist_map = caseloader.load_cart_min_dist(slicei)
dcmstack = caseloader.loadstack(slicei, 'dcm')
ctx, cty = caseloader.loadct(slicei)
aug_polar_patch_batch = np.zeros((10, 256, 256, 3))
aug_polar_cont_batch = np.zeros((10, 256, 2))
offi = 0
mi = 0
while offi < 10 and mi < 50:
mi += 1
if offi == 0:
ofx = 0
ofy = 0
else:
ofx = random.randint(-20, 20)
ofy = random.randint(-20, 20)
cctx = ofx + ctx
ccty = ofy + cty
if min_dist_map[256 + ofx, 256 + ofy] < 0.5:
if mi > 40:
print('mi', mi)
# print('mind',min_dist_map[256+ofx,256+ofy])
continue
# print(ofx,ofy)
new_polar_cont_lumen = tocordpolar(cartcont[0], cctx, ccty)
new_polar_cont_wall = tocordpolar(cartcont[1], cctx, ccty)
aug_polar_cont_batch[offi, :, 0] = new_polar_cont_lumen
aug_polar_cont_batch[offi, :, 1] = new_polar_cont_wall
new_cart_patch = croppatch(dcmstack, ccty, cctx, 256, 256)
new_polar_patch = topolar(new_cart_patch, 256, 256)
aug_polar_patch_batch[offi] = new_polar_patch / np.max(
new_polar_patch)
offi += 1
# aug_polar_patch_batch = caseloader.load_aug_patch(slicei,'polar_patch')
# aug_polar_cont_batch = caseloader.load_aug_patch(slicei,'polar_cont')
for augi in range(len(aug_polar_patch_batch)):
xarray[bi] = aug_polar_patch_batch[augi, :, :, :, None]
yarray[bi] = aug_polar_cont_batch[augi]
bi += 1
if bi == config['batchsize']:
bi = 0
if aug == True:
for offi in random.sample(range(config['height']),
config['rottimes']):
xarray_off = batch_polar_rot(xarray, offi)
yarray_off = batch_polar_rot(yarray, offi)
yield (xarray_off, yarray_off)
else:
yield (xarray, yarray)