def __init__(self, feature_function=None, modelparams={}):
from imcut import pycut
self.working_voxelsize_mm = [1.5, 1.5, 1.5]
self.data = None
self.target = None
# self.cl = sklearn.naive_bayes.GaussianNB()
# self.cl = sklearn.mixture.GMM()
#self.cl = sklearn.tree.DecisionTreeClassifier()
if feature_function is None:
feature_function = intensity_localization_fv
# self.feature_function = feature_function
modelparams_working = {
'fv_type': "fv_extern",
'fv_extern': feature_function,
'type': 'gmmsame',
'params': {
'cvtype': 'full',
'n_components': 15
},
'adaptation': 'original_data',
}
modelparams = add_fv_extern_into_modelparams(modelparams)
modelparams_working.update(modelparams)
self.cl = pycut.Model(modelparams=modelparams_working)
def model_score_from_sliver_data(
# output_file="~/lisa_data/liver_intensity.Model.p",
sliver_reference_dir='~/data/medical/orig/sliver07/training/',
orig_pattern="*orig*[1-9].mhd",
ref_pattern="*seg*[1-9].mhd",
modelparams={},
likelihood_ratio=0.5,
savefig=False,
savefig_fn_prefix='../graphics/bn-symmetry-',
show=False,
label='',
):
"""
:param label: text label added to all records in output table
:param sliver_reference_dir:
:param orig_pattern:
:param ref_pattern:
:param modelparams:
:param likelihood_ratio: float number between 0 and 1, scalar or list. Set the segmentation threshodl
:param savefig:
:param savefig_fn_prefix:
:param show: show images
:return:
"""
import pandas as pd
from imcut import pycut
import sed3
import matplotlib.pyplot as plt
import volumetry_evaluation
sliver_reference_dir = op.expanduser(sliver_reference_dir)
orig_fnames = glob.glob(sliver_reference_dir + orig_pattern)
ref_fnames = glob.glob(sliver_reference_dir + ref_pattern)
orig_fnames.sort()
ref_fnames.sort()
evaluation_all = []
for oname, rname in zip(orig_fnames, ref_fnames):
print(oname)
data3d_orig, metadata = io3d.datareader.read(oname,
dataplus_format=False)
vs_mm1 = metadata['voxelsize_mm']
data3d_seg, metadata = io3d.datareader.read(rname,
dataplus_format=False)
vs_mm = metadata['voxelsize_mm']
mdl = pycut.Model(modelparams=modelparams)
# m0 = mdl.mdl[2]
# len(m0.means_)
vs_mmr = [1.5, 1.5, 1.5]
data3dr = io3d.misc.resize_to_mm(data3d_orig, vs_mm1, vs_mmr)
lik1 = mdl.likelihood_from_image(data3dr, vs_mmr, 0)
lik2 = mdl.likelihood_from_image(data3dr, vs_mmr, 1)
if np.isscalar(likelihood_ratio):
likelihood_ratio = [likelihood_ratio]
for likelihood_ratio_i in likelihood_ratio:
if (likelihood_ratio_i <= 0) or (likelihood_ratio_i >= 1.0):
logger.error("likelihood ratio should be between 0 and 1")
seg = ((likelihood_ratio_i * lik1) >
((1.0 - likelihood_ratio_i) * lik2)).astype(np.uint8)
# seg = (lik1).astype(np.uint8)
seg_orig = io3d.misc.resize_to_shape(seg, data3d_orig.shape)
# seg_orig = io3d.misc.resize_to_shape(seg, data3d_orig.shape)
if show:
plt.figure(figsize=(15, 15))
sed3.show_slices(data3d_orig,
seg_orig,
show=False,
slice_step=20)
# likres = io3d.misc.resize_to_shape(lik1, data3d_orig.shape)
# sed3.show_slices(likres , seg_orig, show=False, slice_step=20)
import re
numeric_label = re.search(".*g(\d+)", oname).group(1)
# plt.figure(figsize = (5,5))
if savefig:
plt.axis('off')
# plt.imshow(symmetry_img)
filename = savefig_fn_prefix + numeric_label + '-lr-' + str(
likelihood_ratio_i) + ".png"
# if np.isscalar(likelihood_ratio):
# filename = filename + ''+str
plt.savefig(filename, bbox_inches='tight')
evaluation = volumetry_evaluation.compare_volumes_sliver(
seg_orig, data3d_seg, vs_mm)
evaluation['likelihood_ratio'] = likelihood_ratio_i
evaluation['numeric_label'] = numeric_label
evaluation['label'] = label
evaluation_all.append(evaluation)
# print evaluation
ev = pd.DataFrame(evaluation_all)
return ev
def automatic_liver_seeds(data3d,
seeds,
voxelsize_mm,
fn_mdl='~/lisa_data/liver_intensity.Model.p',
return_likelihood_difference=True,
gaussian_sigma_mm=[20, 20, 20]):
from imcut import pycut
# fn_mdl = op.expanduser(fn_mdl)
mdl = pycut.Model({
'mdl_stored_file': fn_mdl,
'fv_extern': organ_model.intensity_localization_fv
})
working_voxelsize_mm = np.asarray([1.5, 1.5, 1.5])
gaussian_sigma_mm = np.asarray(gaussian_sigma_mm)
data3dr = imtools.resize_to_mm(data3d, voxelsize_mm, working_voxelsize_mm)
lik1 = mdl.likelihood_from_image(data3dr, working_voxelsize_mm, 0)
lik2 = mdl.likelihood_from_image(data3dr, working_voxelsize_mm, 1)
dl = lik2 - lik1
seeds = add_negative_notrain_seeds(seeds, lik1, lik2)
# Liver seed center
import scipy
# seed tam, kde je to nejpravděpodovnější - moc nefunguje při blbém natrénování
dst = scipy.ndimage.filters.gaussian_filter(dl,
sigma=gaussian_sigma_mm /
working_voxelsize_mm)
# seed1 = np.unravel_index(np.argmax(dl), dl.shape)
# escte jinak
# dáme seed doprostřed oblasti
# dst = scipy.ndimage.morphology.distance_transform_edt(dl>0)
seed1 = np.unravel_index(np.argmax(dst), dst.shape)
# alternativa -
seed1_mm = seed1 * working_voxelsize_mm
print('seed1 ', seed1, ' shape ', dst.shape)
seed1z = seed1[0]
seed1z_mm = seed1_mm[0]
print(seed1z_mm)
add_negative_train_seeds_blobs(dl < 0,
seeds,
working_voxelsize_mm,
voxelsize_mm,
seed1z_mm,
n_seed_blob=3)
seeds = data_manipulation.add_seeds_mm(seeds,
voxelsize_mm, [seed1_mm[0]],
[seed1_mm[1]], [seed1_mm[2]],
label=1,
radius=25,
width=1)
# import sed3
# sed3.show_slices(data3dr, dl > 40.0, slice_step=10)
if return_likelihood_difference:
return seeds, dl
else:
return seeds