def seg_to_small_seg(seg, thres=25, rr=2):
# rr: z/x-y resolution ratio
sz = seg.shape
mask = np.zeros(sz, np.uint8)
for z in np.where(seg.max(axis=1).max(axis=1) > 0)[0]:
tmp = label_cc(seg[z])
ui, uc = np.unique(tmp, return_counts=True)
rl = np.zeros(ui[-1] + 1, np.uint8)
rl[ui[uc < thres]] = 1
rl[0] = 0
mask[z] += rl[tmp]
for y in np.where(seg.max(axis=2).max(axis=0) > 0)[0]:
tmp = label_cc(seg[:, y])
ui, uc = np.unique(tmp, return_counts=True)
rl = np.zeros(ui[-1] + 1, np.uint8)
rl[ui[uc < thres // rr]] = 1
rl[0] = 0
mask[:, y] += rl[tmp]
for x in np.where(seg.max(axis=0).max(axis=0) > 0)[0]:
tmp = label_cc(seg[:, :, x])
ui, uc = np.unique(tmp, return_counts=True)
rl = np.zeros(ui[-1] + 1, np.uint8)
rl[ui[uc < thres // rr]] = 1
rl[0] = 0
mask[:, :, x] += rl[tmp]
return mask
def distance_transform(label: np.ndarray,
bg_value: float = -1.0,
relabel: bool = True,
padding: bool = False,
resolution: Tuple[float] = (1.0, 1.0)):
"""Euclidean distance transform (DT or EDT) for instance masks.
"""
eps = 1e-6
pad_size = 2
if relabel:
label = label_cc(label)
if padding:
# The distance_transform_edt function does not treat image border
# as background. If image border needs to be considered as background
# in distance calculation, set padding to True.
label = np.pad(label, pad_size, mode='constant', constant_values=0)
label_shape = label.shape
all_bg_sample = False
distance = np.zeros(label_shape, dtype=np.float32) + bg_value
semantic = np.zeros(label_shape, dtype=np.uint8)
indices = np.unique(label)
if indices[0] == 0:
if len(indices) > 1: # exclude background
indices = indices[1:]
else: # all-background sample
all_bg_sample = True
if not all_bg_sample:
for idx in indices:
temp1 = label.copy() == idx
temp2 = remove_small_holes(temp1, 16, connectivity=1)
semantic += temp2.astype(np.uint8)
boundary_edt = distance_transform_edt(temp2, resolution)
energy = boundary_edt / (boundary_edt.max() + eps) # normalize
distance = np.maximum(distance, energy * temp2.astype(np.float32))
if padding:
# Unpad the output array to preserve original shape.
distance = array_unpad(distance, get_padsize(
pad_size, ndim=distance.ndim))
semantic = array_unpad(semantic, get_padsize(
pad_size, ndim=distance.ndim))
return distance, semantic
def syn_sem2inst(label: np.array):
# For synaptic polarity, convert semantic annotation to instance
# annotation. It assumes the pre- and post-synaptic masks are
# closely in touch with their parteners.
indices = np.unique(label)
assert list(indices) == [0, 1, 2]
fg = (label != 0).astype(bool)
struct = disk(2, dtype=bool)[np.newaxis, :, :] # only for xy plane
fg = binary_dilation(fg, struct)
segm = label_cc(fg).astype(int)
seg_pos = (label == 1).astype(segm.dtype)
seg_neg = (label == 2).astype(segm.dtype)
seg_pos = seg_pos * (segm * 2 - 1)
seg_neg = seg_neg * (segm * 2)
instance_label = np.maximum(seg_pos, seg_neg)
# Cast the mask to the best dtype to save storage.
max_id = np.amax(np.unique(instance_label))
m_type = getSegType(int(max_id))
return instance_label.astype(m_type)
