def levelset_geo_separete(img,
labels,
niter=10,
prop=-1,
advec=-0.5,
curve=-0.5):
seg = sitk.GetImageFromArray((labels > 0).astype(np.uint8))
init_ls = sitk.SignedMaurerDistanceMap(seg,
insideIsPositive=True,
useImageSpacing=True)
mask = find_boundaries(gvoronoi(label(labels, connectivity=1)),
mode='inner')
img1 = img.copy()
img1[mask] = 0
img_T1 = sitk.GetImageFromArray(img1)
lsFilter = sitk.GeodesicActiveContourLevelSetImageFilter()
lsFilter.SetMaximumRMSError(0.02)
lsFilter.SetNumberOfIterations(niter)
lsFilter.SetAdvectionScaling(advec)
lsFilter.SetCurvatureScaling(curve)
lsFilter.SetPropagationScaling(prop)
# lsFilter.ReverseExpansionDirectionOn()
ls = lsFilter.Execute(init_ls, sitk.Cast(img_T1, sitk.sitkFloat32))
return label(sitk.GetArrayFromImage(ls) > 0, connectivity=1)
def get_labels(self):
temp = label(self.bw.copy(), connectivity=1)
temp = np.zeros(self.bw.shape)
if not self.goodcells:
return label(self.bw)
max_label = temp.max()
for n, i in enumerate(self.all_cells):
for c in i.coords:
temp[c[0], c[1]] = max_label + n + 1
return temp
def search_cut_candidates(self, bw, coords_set):
candidates = []
for n1, (startpt, endpt) in enumerate(coords_set):
# cut and get candidate regionprops
cand_rps = make_candidates_img(bw.copy(), startpt[0], startpt[1],
endpt[0], endpt[1], self.img)
if not len(cand_rps) > len(
np.unique(label(self.cell.filled_image,
connectivity=1))) - 1:
continue
if not all([c.area > np.pi * self.small_rad**2 for c in cand_rps]):
continue
# scoring
for n2, c in enumerate(cand_rps):
# If you get an error here, it might be unresolved error of scikit-image. https://github.com/scikit-image/scikit-image/issues/1470
# edit _regionprops._RegionProperties.moments and moments_central from "astype(np.uint8)" to "astype(np.float)"
if c.minor_axis_length > 1:
c.n1, c.n2 = n1, n2
c.cut_coords = (startpt, endpt)
c.cutline_len = np.sqrt((startpt[0] - endpt[0])**2 +
(startpt[1] - endpt[1])**2)
c.line_total = intensity_below_line(
self.img, startpt[0], startpt[1], endpt[0],
endpt[1]).sum()
candidates.append(c)
return candidates
def levelset_geo(img, labels, advec=3, propagation=0.75, niter=100):
"""Propagate outwards
"""
simg = sitk.GetImageFromArray(img, sitk.sitkFloat32)
labels = label(binary_erosion(labels, np.ones((3, 3))))
seg = sitk.GetImageFromArray(labels)
init_ls = sitk.SignedMaurerDistanceMap(seg, insideIsPositive=True, useImageSpacing=True)
gfil = sitk.GeodesicActiveContourLevelSetImageFilter()
gfil.SetMaximumRMSError(0.02)
gfil.SetNumberOfIterations(niter)
gfil.SetAdvectionScaling(advec)
gfil.SetCurvatureScaling(1)
gfil.SetPropagationScaling(propagation)
gfil.ReverseExpansionDirectionOn()
ls = gfil.Execute(init_ls, sitk.Cast(simg, sitk.sitkFloat32))
ls = label(sitk.GetArrayFromImage(ls) > 0)
return keep_labels(labels, ls)
def run_concave_cut(img, labels, small_rad, large_rad, EDGELEN=6, THRES=180):
store = get_cut_coords(img,
labels,
small_rad,
large_rad,
EDGELEN=EDGELEN,
THRES=THRES)
labels = label(cut_labels(labels, [i for j in store for i in j]),
connectivity=1)
return labels
def levelset_lap(img, labels, NITER=100, CURVE=3, PROP=-1):
simg = sitk.GetImageFromArray(img, sitk.sitkFloat32)
seg = sitk.GetImageFromArray(labels)
init_ls = sitk.SignedMaurerDistanceMap(seg, insideIsPositive=True, useImageSpacing=True)
lsFilter = sitk.LaplacianSegmentationLevelSetImageFilter()
lsFilter.SetMaximumRMSError(0.02)
lsFilter.SetNumberOfIterations(NITER)
lsFilter.SetCurvatureScaling(CURVE)
lsFilter.SetPropagationScaling(PROP)
lsFilter.ReverseExpansionDirectionOn()
ls = lsFilter.Execute(init_ls, sitk.Cast(simg, sitk.sitkFloat32))
ls = label(sitk.GetArrayFromImage(ls) > 0)
return keep_labels(labels, ls)
def convert_labels(lb_ref_to, lb_ref_from, lb_convert):
"""
lb_ref_to:
lb_ref_from:
lb_convert: a labeled image to be converted.
"""
lbmap_to, lbmap_from = zip(*labels_map(lb_ref_to, lb_ref_from))
arr = np.zeros(lb_convert.shape, dtype=np.uint16)
lb = lb_convert.copy()
for n0, n1 in zip(lbmap_to, lbmap_from):
arr[lb == n1] = n0
lb[lb == n1] = 0
remained = label(lb)
remained = remained + arr.max()
remained[remained == arr.max()] = 0
arr = arr + remained
return arr
def dlevel_set(phi, F, niter=20, dt=-0.5, mask=None):
"""
Implementation of Level set method. https://wiseodd.github.io/techblog/2016/11/05/levelset-method/
It has an extra repulsion and masking processes.
niter (int): iteration
dt (float): negative values for propagation, positive values for shrinking
phi (ndarray): labels, set inner objects as -1 and outside as 1.
F (ndarray): img, set boundaries close to 0 and elsewhere close to 1.
mask (ndarray[np.bool]): contours do not cross this line.
"""
if mask is None:
mask = np.zeros(phi.shape, np.bool)
for i in range(niter):
dphi = grad(phi)
dphi_norm = norm(dphi)
region = calc_mask_exclude_overlap(label(phi < 0),
2) # added for absolute repulsion
dphi_norm[region] = 0 # added for absolute repulsion
dphi_norm[mask] = 0 # added constraints based on the image
dphi_t = F * dphi_norm
phi = phi + dt * dphi_t
return phi
def make_candidates_img(cl_label, s0, s1, e0, e1, img):
cut_label = cut_neck(cl_label, s0, s1, e0, e1)
cand_label = label(cut_label, connectivity=1)
cand_rps = regionprops(cand_label, img)
return cand_rps
def num_seg(self):
"""
number of isolated segments
"""
return np.max(label(self.image, connectivity=1))
