def run_vem(data, mask, laplace_init=False):
mu, sigma = moment_matching(data[mask])
S = Segmentation(data, mask=mask, mu=mu, sigma=sigma, beta=0)
S.run(niters=PRE_NITERS)
S = Segmentation(data, mask=mask, mu=S.mu, sigma=S.sigma,
ngb_size=NGB_SIZE, beta=BETA)
e, f, v, vf = [], [], [], []
# Initialize ppm with RW optionally
if laplace_init:
S = run_rw(S)
S.ppm[S.mask] = binarize_ppm(S.ppm[S.mask])
for s in range(len(NITERS)):
# Run VEM for a few iterations
S, es, fs, vs, vfs = _run_vem(S, niters=NITERS[s])
e += es
f += fs
v += vs
vf += vfs
return S, e, f, v, vf
def run_rw(img, mask, mu, sigma):
S = Segmentation(img.get_data(), mask=mask, mu=mu, sigma=sigma,
ngb_size=NGB_SIZE, beta=BETA)
prior, gamma = config_random_walker(S.ext_field(), BETA)
q = random_walker(mask, prior, gamma)
S.ppm[mask] = q
return S
def init_laplace(img, mask, mu, sigma):
S = Segmentation(img.get_data(), mask=mask, mu=mu, sigma=sigma,
ngb_size=NGB_SIZE, beta=BETA)
prior, gamma = config_random_walker(S.ext_field(), BETA)
q = random_walker(mask, prior, gamma)
if BINARIZE:
S.ppm[mask] = binarize_ppm(q)
else:
S.ppm[mask] = q
return S
def init_classical(img, mask, niters=0):
"""
moment matching following by indep EM
"""
mu, sigma = moment_matching(img.get_data()[mask])
S = Segmentation(img.get_data(), mask=mask, mu=mu, sigma=sigma,
beta=0.0)
if niters > 0:
S.run(niters=niters)
S.set_markov_prior(BETA)
return S
def compare_models(filepath):
path, fname = split(filepath)
fid, _ = splitext(fname)
# Input image
img = nb.load(join(path, fid + '.nii'))
# Input mask
mask = img.get_data() > 0
# 4k-model
mu, sigma = moment_matching(img.get_data()[mask], MU4, SIGMA4)
S = Segmentation(img.get_data(), mask=mask, mu=mu, sigma=sigma,
ngb_size=NGB_SIZE, beta=BETA)
S.run(niters=NITERS)
save_map(img, S.ppm, mask, fid, stamp='4K')
save_map(img, S.ppm, mask, fid, stamp='4K_conv', conv_mat=CONV_MAT4)
# 5k-model
mu, sigma = moment_matching(img.get_data()[mask], MU5, SIGMA5)
S = Segmentation(img.get_data(), mask=mask, mu=mu, sigma=sigma,
ngb_size=NGB_SIZE, beta=BETA)
S.run(niters=NITERS)
print S.mu
print S.sigma
save_map(img, S.ppm, mask, fid, stamp='5K')
save_map(img, S.ppm, mask, fid, stamp='5K_conv', conv_mat=CONV_MAT5)
def run_naive(img, mask, mu, sigma, niters=1):
S = Segmentation(img.get_data(), mask=mask, mu=mu, sigma=sigma, beta=0)
S.ve_step()
return S
def run_naive(img, mask, mu, sigma):
S = Segmentation(img.get_data(), mask=mask, mu=mu, sigma=sigma,
ngb_size=NGB_SIZE, beta=BETA)
field = S.ext_field()
S.ppm[mask] = (field.T / field.sum(1)).T
return S
def run_vem(img, mask, mu, sigma):
S = Segmentation(img.get_data(), mask=mask, mu=mu, sigma=sigma,
ngb_size=NGB_SIZE, beta=BETA)
for it in range(NITERS):
S.ve_step()
return S