def relabel_blocks(info,
ids='segm/labels_memb_del',
pf='relabeled',
maxlabel=1):
"""Relabel all blocks in the dataset sequentially."""
for k, v in info.items():
ods = '{}_{}'.format(ids, pf)
print('relabeling block {:03d} (id:{}) to {}'.format(
k, v['postfix'], ods))
# seg = read_image(v, ids=ids, imtype='Label')
fstem = os.path.join(v['datadir'],
'{}_{}'.format(v['base'], v['postfix']))
seg = LabelImage('{}.h5/{}'.format(fstem, ids))
seg.load(load_data=False)
seg = relabel(seg, maxlabel=maxlabel)
if seg.maxlabel != 0:
maxlabel = seg.maxlabel
print('maxlabel = {:08d}'.format(maxlabel))
seg.close()
return maxlabel
def split_segments(seg_path,
ids='/segm/labels_memb_del_relabeled_fix',
outputstem=''):
labels = LabelImage(seg_path)
labels.load(load_data=False)
labels_ds = labels.slice_dataset()
# nuclei
outstem = '{}.h5{}'.format(outputstem, '/nucl/dapi')
maskpath_sauvola = '{}_mask_sauvola'.format(outstem)
maskpath_absmin = '{}_mask_absmin'.format(outstem)
mask_nucl = nucl_mask(maskpath_sauvola, maskpath_absmin)
write(mask_nucl, outstem, '_mask_nuclei', labels, imtype='Mask')
# membranes # TODO: may combine with planarity_mask to make it more data-informed
outstem = '{}.h5{}'.format(outputstem, '/memb/boundary')
mask_memb = memb_mask(labels_ds)
write(mask_memb, outstem, '_mask', labels, imtype='Mask')
outstem = '{}.h5{}'.format(outputstem, ids)
for mask, pf in zip([mask_memb, mask_nucl], ['_memb', '_nucl']):
labs = np.copy(labels_ds)
labs[~mask] = 0
write(labs, outstem, pf, labels, imtype='Label')
def delete_blocks_parallel(image_in, block_idx, postfix='fix'):
im = LabelImage(image_in)
im.load(load_data=False)
vols = {
postfix: 'uint32',
'{}_reseg_mask'.format(postfix): 'bool',
'block_idxs': 'uint16'
}
for pf, dtype in vols.items():
delete_h5_dataset(im, pf=pf)
def copy_blocks_parallel(image_in, block_idx, postfix='fix'):
im = LabelImage(image_in)
im.load(load_data=False)
vols = {
postfix: ['Label', 'uint32'],
'{}_reseg_mask'.format(postfix): ['Mask', 'bool'],
#'{}_peaks'.format(postfix): ['Mask', 'bool'],
'block_idxs': ['Label', 'uint16']
}
#vols = {postfix: ['Mask', 'bool']}
for pf, (imtype, dtype) in vols.items():
copy_h5_dataset(im, imtype, pf, dtype, k=block_idx)
def relabel_parallel(image_in, block_idx, maxlabelfile, pf='relabeled'):
# print('block_idx {:03d} will start from {:10d}'.format(block_idx, maxlabel))
# print('block_idx {:03d}'.format(block_idx))
maxlabels = np.loadtxt(maxlabelfile, dtype=np.uint32)
maxlabel = np.sum(maxlabels[:block_idx])
# print('maxlabel {:10d}'.format(maxlabel))
seg = LabelImage(image_in)
seg.load(load_data=False)
# TODO: into segmentation pipeline: but will probably fail as attr too big
# seg.set_ulabels()
# seg.ds.attrs.create('ulabels', seg.ulabels, dtype='uint32')
relabel(seg, pf, maxlabel)
def read_image(im_info, ids='segm/labels_memb_del', imtype='Label'):
""""Read a h5 dataset as Image object."""
fname = '{}_{}'.format(im_info['base'], im_info['postfix'])
fstem = os.path.join(im_info['datadir'], fname)
if imtype == 'Label':
im = LabelImage('{}.h5/{}'.format(fstem, ids))
elif imtype == 'Mask':
im = MaskImage('{}.h5/{}'.format(fstem, ids))
else:
im = Image('{}.h5/{}'.format(fstem, ids))
im.load(load_data=False)
if imtype == 'Label':
im.set_maxlabel()
return im
def postprocess_features(
seg_paths,
blocksize=[],
blockmargin=[],
blockrange=[],
csv_dir='',
csv_stem='',
feat_pf='_features',
segm_pfs=['full', 'memb', 'nucl'],
ext='csv',
min_size_nucl=50,
save_border_labels=False,
split_features=False,
fset_morph='minimal',
fset_intens='minimal',
):
labels = LabelImage(seg_paths[0], permission='r')
labels.load(load_data=False)
comps = labels.split_path()
csv_dir = csv_dir or comps['dir']
mpi = wmeMPI(usempi=False)
mpi.set_blocks(labels, blocksize, blockmargin, blockrange)
mpi.scatter_series()
li = []
for i in mpi.series:
print('processing block {:03d}'.format(i))
block = mpi.blocks[i]
# read the csv's
filestem = '{}_{}{}'.format(csv_stem, block['id'], feat_pf)
dfs = {}
for segm_pf in segm_pfs:
filename = '{}_{}.{}'.format(filestem, segm_pf, ext)
filepath = os.path.join(csv_dir, filename)
dfs[segm_pf] = pd.read_csv(filepath, index_col='label', header=0)
if len(dfs['full'].index) == 0:
continue
# select features
# metrics=['mean', 'median', 'variance', 'min', 'max']
metrics = ['mean']
feat_names = get_feature_names(fset_morph, fset_intens, metrics)
df = select_features(dfs, feat_names, min_size_nucl, split_features)
#df = rename_columns(df, metrics=metrics)
# label filtering: only select full segments
filestem = '{}_{}'.format(csv_stem, block['id'])
sl_path = os.path.join(csv_dir,
'{}_smalllabels.pickle'.format(filestem))
with open(sl_path, 'rb') as f:
slabels = pickle.load(f)
bl_path = os.path.join(csv_dir,
'{}_borderlabels.pickle'.format(filestem))
if save_border_labels:
labels.slices = block['slices']
blabels = find_border_labels(labels)
with open(bl_path, 'wb') as f:
pickle.dump(blabels, f)
else:
with open(bl_path, 'rb') as f:
blabels = pickle.load(f)
blabels -= slabels
df = df.drop(labels=blabels)
li.append(df)
combined_df = pd.concat(li, keys=mpi.series)
combined_df.index.names = ['block', 'label']
combined_df.reset_index(inplace=True)
combined_df.drop_duplicates(subset='label', inplace=True)
outputpath = os.path.join(csv_dir, '{}{}.csv'.format(csv_stem, feat_pf))
combined_df.to_csv(outputpath, index=True, encoding='utf-8-sig')
return combined_df
def export_regionprops(
seg_paths,
seg_names=['full', 'memb', 'nucl'],
data_paths=[],
data_names=[],
aux_data_path=[],
downsample_factors=[1, 1, 1],
outputstem='',
blocksize=[],
blockmargin=[],
blockrange=[],
channels=[],
filter_borderlabels=False,
min_labelsize=0,
split_features=False,
fset_morph=['label'],
fset_intens=['mean_intensity'],
fset_addit=['com_z', 'com_y', 'com_x'],
):
# load the segments: ['full'] or ['full', 'memb', 'nucl']
label_ims = {}
pfs = seg_names[:len(seg_paths)]
for pf, seg_path in zip(pfs, seg_paths):
im = LabelImage(seg_path, permission='r')
im.load(load_data=False)
label_ims[pf] = im
comps = label_ims['full'].split_path()
# prepare parallel processing
mpi_label = wmeMPI(usempi=False)
blocksize = blocksize or label_ims['full'].dims
mpi_label.set_blocks(label_ims['full'], blocksize, blockmargin, blockrange)
mpi_label.scatter_series()
# load the data
data_ims = {}
mpi_data = wmeMPI(usempi=False)
for i, data_path in enumerate(data_paths):
pf = 'im{:02d}'.format(i)
data = Image(data_path, permission='r')
data.load(load_data=False)
ch_idx = data.axlab.index('c')
# FIXME channels for multiple data_paths
chs = channels or [ch for ch in range(data.dims[ch_idx])]
names = [data_names.pop(0) for _ in range(len(chs))]
data_ims[pf] = {'im': data, 'ch': chs, 'names': names}
""" TODO
try:
mpi_data.blocks = [
{'id': split_filename(comps['file'])[0]['postfix'],
'slices': dset_name2slices(comps['file'], axlab=data.axlab, shape=data.dims),
'path': '',},
]
except:
"""
mpi_data.set_blocks(data, blocksize, blockmargin, blockrange)
border_labelset = set([])
# if filter_borderlabels:
# outstem = outputstem or label_ims['full'].split_path()['base']
# outstem += '_dataset'
# border_labelset |= filter_borders(label_ims['full'], outstem)
dfs = []
for i in mpi_label.series:
print('processing block {:03d} with id: {}'.format(
i, mpi_label.blocks[i]['id']))
dfs.append(
process_block(
mpi_label.blocks[i],
mpi_data.blocks[i],
label_ims,
split_features,
data_ims,
min_labelsize,
channels,
filter_borderlabels,
fset_morph,
fset_intens,
fset_addit,
border_labelset,
outputstem,
aux_data_path,
downsample_factors,
))
return dfs
def csv_to_im(
image_in,
csv_path,
labelkey='label',
key='dapi',
name='',
maxlabel=0,
normalize=False,
scale_uint16=False,
replace_nan=False,
channel=-1,
outpath='',
):
"""Write segment backprojection."""
if isinstance(image_in, Image):
labels = image_in
else:
labels = LabelImage(image_in)
labels.load(load_data=False)
if not maxlabel:
labels.set_maxlabel()
maxlabel = labels.maxlabel
if csv_path.endswith('.csv'):
df = pd.read_csv(csv_path)
df = df.astype({labelkey: int})
elif csv_path.endswith('.h5ad'):
import scanpy as sc
adata = sc.read(csv_path)
if not csv_path.endswith('_nofilter.h5ad'):
adata.X = adata.raw.X
df = adata.obs[labelkey].astype(int)
df = pd.concat([df, adata[:, key].to_df()], axis=1)
# for key in keys: # TODO
fw = np.zeros(maxlabel + 1, dtype='float')
for index, row in df.iterrows():
fw[int(row[labelkey])] = row[key]
if replace_nan:
fw = np.nan_to_num(fw)
if normalize:
def normalize_data(data):
"""Normalize data between 0 and 1."""
data = data.astype('float64')
datamin = np.amin(data)
datamax = np.amax(data)
data -= datamin
data *= 1 / (datamax - datamin)
return data, [datamin, datamax]
fw_n, fw_minmax = normalize_data(fw)
fw_n *= 65535
fw = fw_n
elif scale_uint16: # for e.g. pseudotime / FA / etc / any [0, 1] vars
fw *= 65535
out = labels.forward_map(list(fw))
if outpath.endswith('.ims'):
mo = Image(outpath, permission='r+')
mo.load(load_data=False)
if channel >= 0 and channel < mo.dims[3]:
ch = channel
else:
mo.create()
ch = mo.dims[3] - 1
mo.slices[3] = slice(ch, ch + 1, 1)
mo.write(out.astype(mo.dtype)) # FIXME: >65535 wraps around
cpath = 'DataSetInfo/Channel {}'.format(ch)
name = name or key
mo.file[cpath].attrs['Name'] = np.array([c for c in name], dtype='|S1')
mo.close()
elif outpath.endswith('.nii.gz'):
props = labels.get_props()
if not labels.path.endswith('.nii.gz'):
props = transpose_props(props, outlayout='xyz')
out = out.transpose()
mo = write_output(outpath, out, props)
else:
outpath = outpath or gen_outpath(labels, key)
mo = write_output(outpath, out, labels.get_props())
return mo