def lab2bgdist(lab):
distimg = lab.copy()
distimg[lab != 0] = 1
bor = label_tools.find_boundaries(lab)
distimg[bor] = 0 ## mem is also bg!
distimg = distance_transform_edt(distimg)
hx = np.array([1, 1, 1]) / 3
distimg = gputools.convolve_sep3(distimg, hx, hx, hx, sub_blocks=(1, 1, 1))
distimg = gputools.convolve_sep3(distimg, hx, hx, hx, sub_blocks=(1, 1, 1))
distimg[lab == 0] = 0
distimg[bor] = 0
for i in range(1, lab.max()):
m = lab == i
distimg[m] = distimg[m] / max(distimg[m].max(), 1)
return distimg
def test_conv_sep3_numpy():
Nz, Nx, Ny = 128, 203, 303
d = np.zeros((Nz, Ny, Nx), np.float32)
d[::10, ::10, ::10] = 1.
hx = np.ones(8)
hy = np.ones(3)
hz = np.ones(11)
res1 = gputools.convolve_sep3(d, hx, hy, hz, sub_blocks=(1, 1, 1))
res2 = gputools.convolve_sep3(d, hx, hy, hz, sub_blocks=(7, 4, 3))
assert np.allclose(res1, res2)
return res1, res2
def apply(self, data):
sigs = (self.sx, self.sy, self.sz)
Ns = [2 * s + 1 for s in sigs]
xs = [np.arange(-N, N + 1) for N in Ns]
hs = [np.exp(-x**2 / 2. / s**2) for x, s in zip(xs, sigs)]
hs = [1. * h / sum(h) for h in hs]
return gputools.convolve_sep3(data, *hs)
def apply(self,data):
sigs = (self.sx,self.sy,self.sz)
Ns = [2*s+1 for s in sigs]
xs = [np.arange(-N,N+1) for N in Ns]
hs = [np.exp(-x**2/2./s**2) for x,s in zip(xs,sigs)]
hs = [1.*h/sum(h) for h in hs]
return gputools.convolve_sep3(data, *hs)
def test_conv_gpu():
N = 128
d = np.zeros((N, N + 3, N + 5), np.float32)
d[N // 2, N // 2, N // 2] = 1.
h = np.exp(-10 * np.linspace(-1, 1, 17)**2)
res = gputools.convolve_sep3(d, h, h, h)
def test_conv_sep3_numpy():
Nz, Nx, Ny = 128, 203, 303
d = np.zeros((Nz, Ny,Nx),np.float32)
d[::10,::10,::10] = 1.
hx = np.ones(8)
hy = np.ones(3)
hz = np.ones(11)
res1 = gputools.convolve_sep3(d,hx,hy,hz, sub_blocks=(1,1,1))
res2 = gputools.convolve_sep3(d,hx,hy,hz, sub_blocks=(7,4,3))
assert np.allclose(res1,res2)
return res1, res2
def test_conv_gpu():
N = 128
d = np.zeros((N,N+3,N+5),np.float32)
d[N/2,N/2,N/2] = 1.
h = np.exp(-10*np.linspace(-1,1,17)**2)
res = gputools.convolve_sep3(d,h,h,h)
def segment(d, params):
img2 = gputools.denoise.nlm3(d['img'][..., 0], sigma=params['sigma'])
hx = np.array([1, 1, 1]) / 3
for _ in range(params['nblur']):
img2 = gputools.convolve_sep3(img2, hx, hx, hx)
hyp = watershed(-img2,
label(d['cen'] > 0)[0],
mask=img2 > img2.mean(
(1, 2), keepdims=True) * params['m_mask'],
compactness=params['compactness'])
return hyp
def segment_pimg_img(d, params):
img = d['img'][..., 0].astype(np.float32)
pimg = d['pimg']
hx = np.array([1, 1, 1]) / 3
img2 = gputools.denoise.nlm3(img, sigma=params['sigma'])
for _ in range(params['nblur']):
img2 = gputools.convolve_sep3(img2, hx, hx, hx)
hyp = watershed(-img2,
label(pimg > params['thi'])[0],
mask=img2 > params['fmask'](img2) * params['m_mask'],
compactness=params['compactness'])
return hyp
def test_simple3():
d = np.zeros((100, 100, 100))
d[50, 50, 50] = 1.
hx = np.random.uniform(0, 1, 7)
hy = np.random.uniform(0, 1, 11)
hz = np.random.uniform(0, 1, 17)
# hy = np.ones(1)
# hz = np.ones(1)
res1 = _conv_sep3_numpy(d, hx, hy, hz)
res2 = gputools.convolve_sep3(d, hx, hy, hz)
return res1, res2
def segment_otsu(d, params):
img2 = gputools.denoise.nlm3(d['img'][..., 0], sigma=params['sigma'])
hx = np.array([1, 1, 1]) / 3
for _ in range(params['nblur']):
img2 = gputools.convolve_sep3(img2, hx, hx, hx)
img2 = norm_szyxc(img2, (0, 1, 2))
# local_otsu = np.array([rank.otsu(img2[i], disk(params['m_mask']))/255 for i in range(img2.shape[0])])
otsu = otsu3d(img2)
hyp = watershed(-img2,
label(d['cen'] > 0)[0],
mask=img2 > otsu,
compactness=params['compactness'])
return hyp
def segment_and_plot(img, with_border_cells=True):
potential = 1 - img[..., 1] + img[..., 0]
potential = np.clip(potential, 0, 1)
hx = gaussian(21, 2.0)
potential = gputools.convolve_sep3(potential, hx, hx, hx)
potential = lib.normalize_percentile_to01(potential, 0, 100)
# nhl,hyp = watershed(potential, c1=0.5, c2=0.9)
hyp = watershed(potential, label(potential < 0.1)[0], mask=potential < 0.5)
if with_border_cells == False:
mask_borders = lib.mask_border_objs(hyp)
hyp[mask_borders] = 0
nhl = lib.hyp2nhl(hyp, img[..., 1])
nhl = np.array(nhl)
areas = np.array([n['area'] for n in nhl])
xmom = np.array([n['moments_img'][0, 0, 0] / n['area'] for n in nhl])
col = plt.scatter(np.log2(areas), xmom) #np.log2(xmom))
# selector = view.SelectFromCollection(plt.gca(), col)
return nhl, hyp
def build_rawdata2(homedir):
homedir = Path(homedir)
## load data
# img = imread(str(homedir / 'data/img006.tif'))
img = np.load(str(homedir / 'data/img006_noconv.npy'))
img = img[1]
img = perm(img, "ZCYX", "ZYXC")
img = norm_szyxc_per(img, (0, 1, 2))
imgsem = {'axes': "ZYXC", 'nuc': 1, 'mem': 0, 'n_channels': 2}
points = lib.mkpoints()
cen = np.zeros(img.shape[:-1])
cen[list(points.T)] = 1
x = img[..., 1]
hx = np.array([1, 1, 1]) / 3
x = gputools.convolve_sep3(x, hx, hx, hx)
lab = watershed(-x, label(cen)[0], mask=x > x.mean())
bor = label_tools.find_boundaries(lab)
bor = np.array([morph.binary_dilation(b) for b in bor]) ## just x,y
bor = np.array([morph.binary_dilation(b) for b in bor]) ## just x,y
# bor = np.array([morph.binary_dilation(b) for b in bor]) ## just x,yp
# bor = morph.binary_erosion(bor)
# bor = morph.binary_dilation(bor)
# bor = morph.binary_erosion(bor)
# bor = morph.binary_dilation(bor)
# bor = morph.binary_dilation(bor)
lab[lab != 0] = 1
lab[bor] = 2
labsem = {'n_classes': 3, 'nuc': 1, 'mem': 2, 'bg': 0, 'axes': 'ZYX'}
res = dict()
res['img'] = img
res['imgsem'] = imgsem
res['lab'] = lab
res['labsem'] = labsem
res['points'] = points
res['cen'] = cen
# res['mask_labeled_slices'] = mask_labeled_slices
# res['inds_labeled_slices'] = inds_labeled_slices
# res['gt_slices'] = gt_slices
return res
def _convolve_rand(dshape, hshapes):
print("convolving test: dshape = %s, hshapes = %s" % (dshape, hshapes))
np.random.seed(1)
d = np.random.uniform(0, 1, dshape).astype(np.float32)
hs = [
np.random.uniform(1, 2, hshape).astype(np.float32)
for hshape in hshapes
]
if len(hs) == 2:
out1 = _conv_sep2_numpy(d, *hs)
out2 = gputools.convolve_sep2(d, *hs)
elif len(hs) == 3:
out1 = _conv_sep3_numpy(d, *hs)
out2 = gputools.convolve_sep3(d, *hs)
print("difference: ", np.amax(np.abs(out1 - out2)))
npt.assert_allclose(out1, out2, rtol=1.e-2, atol=1.e-5)
return out1, out2
def apply(self,data):
N = 2*self.sigma+1
x = np.arange(-N,N+1)
h = np.exp(-x**2/2./self.sigma**2)
h *= 1./sum(h)
return gputools.convolve_sep3(data, h, h, h)
def apply(self, data):
N = 2 * self.sigma + 1
x = np.arange(-N, N + 1)
h = np.exp(-x**2 / 2. / self.sigma**2)
h *= 1. / sum(h)
return gputools.convolve_sep3(data, h, h, h)
def build_rawdata(homedir):
homedir = Path(homedir)
## load data
# img = imread(str(homedir / 'data/img006.tif'))
img = np.load(str(homedir / 'data/img006_noconv.npy'))
img = img[1]
img = perm(img, "ZCYX", "ZYXC")
img = norm_szyxc_per(img, (0, 1, 2))
img = img.astype(np.float32)
points = lib.mkpoints()
imgsem = {'axes': "ZYXC", 'nuc': 1, 'mem': 0, 'n_channels': 2}
# img = img[:70,::2,::2]
scale = (2.5, .5, .5)
def f(idx):
return gputools.scale(img[..., idx[0]], scale)
img = ts.broadcast_over(f, (2, ))
img = perm(img, "czyx", "zyxc")
points = (points * scale).astype(np.int)
# points[:,0] = points[:,0].clip(0,img.shape[0]-1)
# scale!
# img = zoom(img, (5,1,1,1))
if False:
shape = np.array(img.shape)
res = patchmaker.patchtool({
'img': shape,
'patch': shape // 2,
'grid': (2, 2, 2, 2)
})
scaled = [
gputools.scale((img[ss][..., 0]).astype(np.float32), (5, 1, 1))
for ss in res['slices']
]
ends = res['starts'] + shape // 2 * [5, 1, 1, 1]
container = np.zeros(shape * [5, 1, 1, 1])
big_patches = patchmaker.starts_ends_to_slices(res['starts'], ends)
for i in range(len(big_patches)):
container[big_patches[i]] = scaled[i][..., np.newaxis]
container = img.copy()
cen = np.zeros(container.shape[:-1])
cen[list(points.T)] = 1
x = container[..., 1]
hx = np.array([1, 1, 1]) / 3
x = gputools.convolve_sep3(x, hx, hx, hx, sub_blocks=(1, 1, 1))
lab = watershed(-x, label(cen)[0], mask=x > x.mean())
# bor = np.array([morph.binary_dilation(b) for b in bor]) ## just x,y
# bor = np.array([morph.binary_dilation(b) for b in bor]) ## just x,y
if False:
target = lab[::2, ::2, ::2]
points = points // 2
bg = target == 0
fg = target == 1
target[fg] = 0.0
hx = np.array([1, 1, 1]) / 3
for i in range(200):
target[bg] = 0
target[list(points.T)] = 1
target = gputools.convolve_sep3(target,
hx,
hx,
hx,
sub_blocks=(2, 2, 2))
points = points * 2
target = gputools.scale(target, (2, 2, 2))
if False:
xx = np.linspace(-2, 2, 80)
xx = np.exp(-xx**2 / 2)
xx = xx / xx.sum()
gauss = gputools.convolve_sep3(cen, xx, xx, xx, sub_blocks=(2, 2, 2))
if True:
target = lab2bgdist(lab)
if False:
target = target * gauss
target = np.exp(-target / 10)
res = dict()
res['source'] = container
res['imgsem'] = imgsem
res['target'] = target
res['points'] = points
res['cen'] = cen
return res
'NET (border weights)', 'NET (t007)'
]
data = [ys_ilastik, ys_t001, ys_t002, ys_t005, ys_t006, ys_t007]
plt.figure()
for i in range(6):
name = names[i]
img = data[i]
# if i==5:
# img = img[1::5]
# potential = 1 - img[...,1] + img[...,0]
# potential = lib.normalize_percentile_to01(potential, 0, 100)*2
# hyp = watershed(potential, seeds, mask=potential<0.5)
# nhl = lib.hyp2nhl(hyp)
potential = img[..., 1] # - img[...,0]
hx = gaussian(21, 2.0)
potential = gputools.convolve_sep3(potential, hx, hx, hx)
potential = lib.normalize_percentile_to01(potential, 0, 100)
nhl, hyp = lib.two_var_thresh(potential, c1=0.5, c2=0.9)
areas = np.array([n['area'] for n in nhl])
if i == 5:
areas = areas / 5
plt.plot(sorted(np.log2(areas)), label=name)
plt.legend()
## Let's take a close look at segmentations using slice views
data = [ys_ilastik, ys_t001, ys_t007]
a, b, c, d = ys_t001.shape
bigimg = np.zeros((a, b, 3 * c), dtype=np.float)
for i in range(3):
img6_t1 = img6[1, ..., 1].copy()
