def load_sample(self, lq, hq):
self.sample_files = [lq, hq]
lq_oct = io.OCTVolume(lq,
load=True,
pad={},
normalize={'min': 'mean-0.5'})
hq_oct = io.OCTVolume(hq,
load=True,
pad={},
normalize={'min': 'mean-0.5'})
hq = np.reshape(hq_oct.image_data, (49, 512, 512, 1))
lq = np.reshape(lq_oct.image_data, (49, 512, 512, 1))
self.samples = {'hq': hq, 'lq': lq}
def load(path):
print('loading: ' + path)
try:
oct = oct_io.OCTVolume(path, load=True, pad={}, normalize=True)
except InvalidDicomError:
print('invalid file:' + path)
return None
return oct.image_data
def setup():
# load model
checkpoint = '/media/network/DL_PC/ilja/cycoct-skip_processed/gen_norm-in_act-selu_scale-3_res-6x3_f-16/dis_norm-id_act-selu_f-1/10-cyc_1-dis/cGAN.ckpt-333200'
garch = models.CyclicGAN.generate_gen_arch(3,
6,
3,
tf.nn.selu,
filters=16,
skip_conn=True)
darch = models.CyclicGAN.generate_dis_arch(512,
tf.nn.selu,
norm_func=models.identity)
logdir = '/home/kazuki/testing'
inputs = {
'hq': tf.placeholder(tf.float32, (1, 512, 512, 1), 'HQ-Input'),
'lq': tf.placeholder(tf.float32, (1, 512, 512, 1), 'LQ-Input')
}
graph = tf.get_default_graph()
cgan = models.CyclicGAN(garch,
darch,
inputs,
None,
logdir,
graph=graph,
inference_mode=True)
# load in sample dicoms
# lq_oct = io.OCTVolume('../DeepOCTPrior/test_octs/lo.dcm')
# hq_oct = io.OCTVolume('../DeepOCTPrior/test_octs/hi.dcm')
lq_oct = io.OCTVolume(
'/media/network/ImageRepository/cimg_db/24573/564875/6/R/pat.dcm')
hq_oct = io.OCTVolume(
'/media/network/ImageRepository/cimg_db/24573/564879/6/R/pat.dcm')
lq_oct.load()
hq_oct.load()
hq_oct.pad()
lq_oct.pad()
return hq_oct, lq_oct, cgan, checkpoint, inputs
np.random.shuffle(indices_hq)
indices_hq = list(indices_hq)
# set directories for export
export_dir_lq = os.path.join(export_dir, 'LQ-Comparison')
export_dir_hq = os.path.join(export_dir, 'HQ-Comparison')
# initialize record of exported slices
export_record = []
for (pid, pos), frame in file_frame.groupby(level=[0, 1]):
# load OCTs
lq_volume = oct_io.OCTVolume(frame.loc[pid, pos, 'low']['Filename'],
load=True,
pad={},
normalize={
'min': 'mean-0.5'
}).image_data
hq_volume = oct_io.OCTVolume(frame.loc[pid, pos, 'high']['Filename'],
load=True,
pad={},
normalize={
'min': 'mean-0.5'
}).image_data
for lq, hq in zip(lq_volume, hq_volume):
# enhance image
enhanced = model.generate(lq[np.newaxis, ..., np.newaxis],
'hq')[0, ..., 0]
def __call__(self, csv):
print('beginning analysis...',
'\tcsv:\t{}'.format(csv), sep='\n')
# set up input dataframe with data and filenames
data = pd.DataFrame.from_csv(csv, sep=',', index_col=None)
data['Filename'] = data.apply(partial(build_filenames, root=self.image_repository, columns=self.path_columns), axis=1)
data.set_index(['PID', 'Position', 'Quality'], inplace=True)
total_slices = data.loc[:, 'ImgCount'].sum() / 4
analized_slices = 0
print('\tnumber of datapoins:\t{}'.format(len(data.index)))
# initialize result list for accumulating measurements
results = []
# cycle over volumes in test set
for (pid, pos), frame in data.groupby(level=[0, 1]):
# prepare template for results
result_template = {'PID': pid,
'SID': frame.loc[pid, pos, 'low']['SID'],
'Slice': None,
'Position': pos,
'ExamType1': frame.loc[pid, pos, 'low']['ExamType1'],
'Filename': frame.loc[pid, pos, 'low']['Filename'],
'Method': None}
# load OCTs
lq = io.OCTVolume(frame.loc[pid, pos, 'low']['Filename'], load=True, pad={}, normalize=True).image_data
hq = io.OCTVolume(frame.loc[pid, pos, 'high']['Filename'], load=True, pad={}, normalize=True).image_data
# register volumes
lq = self.registrator.register_volume(hq, lq, offsets=0.5)
# cycle over b-scans in volume
for index, (image, reference) in enumerate(zip(lq, hq)):
# skip failed registrations
if image is None:
analized_slices += 1
continue
rois = self.get_rois(reference)
background = self.get_background(reference)
# initialize dict for the denoised images
datapoints = [Datapoint(template=result_template, index=index, reference=reference,
rois=rois, background=background, method='original', image=image)]
# generate all denoised images
# self.denoising = partial(self.apply_denoising, datapoint=datapoints[0])
# datapoints += self.pool.map(self.denoising, self.methods.keys())
datapoints += [self.apply_denoising(datapoints[0], method) for method in self.methods.keys()]
# perform measurements on all denoised images
# results += self.pool.map(self.apply_measurements, datapoints)
results += [self.apply_measurements(datapoint) for datapoint in datapoints]
analized_slices += 1
print('\r\tprogress: \t{}%'.format(round(100*analized_slices/total_slices, 2)))
# convert results to a dataframe and save
results_frame = pd.DataFrame(data=results)
results_frame.to_csv(SAVEFILE)
return results_frame