def run_lap(img0, img1, labels0, labels1, DISPLACEMENT=30, MASSTHRES=0.2):
'''Linear assignment problem for mammalian cells.
Cost matrix is simply the distance.
costDie and costBorn are variables changing over frame. Update it through holder.
Args:
DISPLACEMENT (int): The maximum distance (in pixel)
MASSTHRES (float): The maximum difference of total intensity changes.
0.2 means it allows for 20% total intensity changes.
'''
labels = -labels1.copy()
rps0 = regionprops(labels0, img0)
rps1 = regionprops(labels1, img1)
if not rps0 or not rps1:
return labels0, labels
dist = cdist([i.centroid for i in rps0], [i.centroid for i in rps1])
massdiff = calc_massdiff(rps0, rps1)
'''search radius is now simply set by maximum displacement possible.
In the future, I might add data-driven approach mentioned in LAP paper (supple pg.7)'''
dist[dist >
DISPLACEMENT] = np.Inf # assign a large cost for unlikely a pair
# dist[abs(massdiff) > MASSTHRES] = np.Inf
cost = dist
if cost.shape[0] == 0 or cost.shape[1] == 0:
return labels0, labels
# Define initial costBorn and costDie in the first frame
if not hasattr(holder, 'cost_born') or not hasattr(holder, 'cost_die'):
holder.cost_born = np.percentile(cost[~np.isinf(cost)], 80)
holder.cost_die = np.percentile(cost[~np.isinf(cost)], 80)
# try:
binary_cost = call_lap(cost, holder.cost_die, holder.cost_born)
# The first assignment of np.Inf is to reduce calculation of linear assignment.
# This part will make sure that cells outside of these range do not get connected.
binary_cost[(np.abs(massdiff) > MASSTHRES)] = False
binary_cost[(dist > DISPLACEMENT)] = False
gp, gc = np.where(binary_cost)
idx0, idx1 = list(gp), list(gc)
for i0, i1 in zip(idx0, idx1):
labels[labels1 == rps1[i1].label] = rps0[i0].label
labels0[labels0 == rps0[i0].label] = -rps0[i0].label
# update cost
linked_dist = [dist[i0, i1] for i0, i1 in zip(idx0, idx1)]
if linked_dist:
cost = np.max(linked_dist) * 1.05
if cost != 0: # solver freezes if cost is 0
holder.cost_born, holder.cost_die = cost, cost
return labels0, labels
def caller(inputs_list, inputs_labels_list, output, primary, secondary):
make_dirs(dirname(abspath(output)))
inputs_list = [inputs_list, ] if isinstance(inputs_list[0], str) else inputs_list
inputs_labels_list = [inputs_labels_list, ] if isinstance(inputs_labels_list[0], str) else inputs_labels_list
obj_names = [basename(dirname(i[0])) for i in inputs_labels_list] if primary is None else primary
ch_names = [basename(dirname(i[0])) for i in inputs_list] if secondary is None else secondary
store = []
for inputs, ch in zip(inputs_list, ch_names):
for inputs_labels, obj in zip(inputs_labels_list, obj_names):
logger.info("Channel {0}: {1} applied...".format(ch, obj))
for frame, (path, pathl) in enumerate(zip(inputs, inputs_labels)):
img, labels = imread(path), lbread(pathl, nonneg=False)
cells = regionprops(labels, img)
if (labels < 0).any():
cells = add_parent(cells, labels)
[setattr(cell, 'frame', frame) for cell in cells]
cells = [Cell(cell) for cell in cells]
store.append(cells)
logger.info("\tmaking dataframe...")
df = multi_index([i for ii in store for i in ii], obj, ch)
if exists(join(output, 'df.csv')):
ex_df = pd.read_csv(join(output, 'df.csv'), index_col=['object', 'ch', 'prop', 'frame'])
ex_df.columns = pd.to_numeric(ex_df.columns)
ex_df = ex_df.astype(np.float32)
df = pd.concat([df, ex_df])
df.to_csv(join(output, 'df.csv'))
larr = df2larr(df)
larr.save(join(output, 'df.npz'))
logger.info("\tdf.npz saved.")
def caller(inputs, inputs_labels, output, functions, params):
make_dirs(output)
# Make cells. cells are a list of regionproperties or subclasses.
logger.info('Postprocess.\tcollecting cells...')
store = []
for frame, (path, pathl) in enumerate(zip(inputs, inputs_labels)):
img, labels = imread(path), lbread(pathl)
cells = regionprops(labels, img)
cells = [Cell(cell) for cell in cells]
for cell in cells:
cell.frame = frame
if frame > 0:
all_labels = [i.label for i in store[frame - 1]]
if cell.label in all_labels:
store[frame - 1][all_labels.index(cell.label)].nxt = cell
store.append(cells)
cells = [i for j in store for i in j]
# Each function receives cells (regionprops) and finally return labels generated by cells.label
for function, param in zip(functions, params):
logger.info('\trunning {0}'.format(function))
func = getattr(postprocess_operation, function)
cells = func(cells, **param)
logger.info('\tsaving images...')
for frame, (path, pathl) in enumerate(zip(inputs, inputs_labels)):
labels = cells2labels(cells, frame, lbread(pathl))
imsave(labels, output, path, dtype=np.int16)
def nn_closer(img0, img1, labels0, labels1, DISPLACEMENT=30, MASSTHRES=0.25):
labels = -labels1.copy()
rps0 = regionprops(labels0, img0)
rps1 = regionprops(labels1, img1)
if not rps0 or not rps1:
return labels0, labels
dist = cdist([i.centroid for i in rps0], [i.centroid for i in rps1])
massdiff = calc_massdiff(rps0, rps1)
binary_cost = (dist < DISPLACEMENT) * (abs(massdiff) < MASSTHRES)
binary_cost = pick_closer_cost(binary_cost, dist)
binary_cost = pick_closer_cost(binary_cost.T, dist.T).T
idx1, idx0 = find_one_to_one_assign(binary_cost)
for i0, i1 in zip(idx0, idx1):
labels[labels1 == rps1[i1].label] = rps0[i0].label
labels0[labels0 == rps0[i0].label] = -rps0[i0].label
return labels0, labels
def caller(inputs_list, inputs_labels_list, output, primary, secondary):
make_dirs(dirname(abspath(output)))
inputs_list = [
inputs_list,
] if isinstance(inputs_list[0], str) else inputs_list
inputs_labels_list = [
inputs_labels_list,
] if isinstance(inputs_labels_list[0], str) else inputs_labels_list
obj_names = [basename(dirname(i[0]))
for i in inputs_labels_list] if primary is None else primary
ch_names = [basename(dirname(i[0]))
for i in inputs_list] if secondary is None else secondary
for inputs, ch in zip(inputs_list, ch_names):
for inputs_labels, obj in zip(inputs_labels_list, obj_names):
logger.info("Channel {0}: {1} applied...".format(ch, obj))
arr = np.ones((MAX_NUMCELL, len(PROP_SAVE), len(inputs)),
np.float32) * np.nan
for frame, (path, pathl) in enumerate(zip(inputs, inputs_labels)):
img, labels = imread(path), lbread(pathl, nonneg=False)
cells = regionprops(labels, img)
if (labels < 0).any():
cells = add_parent(cells, labels)
[setattr(cell, 'frame', frame) for cell in cells]
cells = [Cell(cell) for cell in cells]
tarr = _cells2array(cells)
index = tarr[:, 1].astype(np.int32)
arr[index, :, frame] = tarr
logger.info("\tmaking dataframe...")
cellids = np.where(~np.isnan(arr[:, 0, :]).all(axis=1))[0]
marr = np.zeros((len(cellids), arr.shape[1], arr.shape[2]))
for pn, i in enumerate(cellids):
marr[pn] = arr[i]
sarr = np.swapaxes(marr, 0, 2)
narr = sarr.reshape((sarr.shape[0] * sarr.shape[1], sarr.shape[2]),
order='F')
index = pd.MultiIndex.from_product(
[obj, ch, PROP_SAVE, range(arr.shape[-1])],
names=['object', 'ch', 'prop', 'frame'])
df = pd.DataFrame(narr, index=index, columns=cellids)
if exists(join(output, FILE_NAME + '.csv')):
ex_df = pd.read_csv(
join(output, FILE_NAME + '.csv'),
index_col=['object', 'ch', 'prop', 'frame'])
ex_df.columns = pd.to_numeric(ex_df.columns)
ex_df = ex_df.astype(np.float32)
df = pd.concat([df, ex_df])
df.to_csv(join(output, FILE_NAME + '.csv'))
larr = df2larr(df)
larr.save(join(output, FILE_NAME + '.npz'))
logger.info("\t" + FILE_NAME + ".npz saved.")
def nearest_neighbor(img0,
img1,
labels0,
labels1,
DISPLACEMENT=20,
MASSTHRES=0.2):
"""
labels0 and labels1: the positive values for non-tracked objects and the negative values for tracked objects.
"""
labels = -labels1.copy()
rps0 = regionprops(labels0, img0)
rps1 = regionprops(labels1, img1)
if not rps0 or not rps1:
return labels0, labels
dist = cdist([i.centroid for i in rps0], [i.centroid for i in rps1])
massdiff = calc_massdiff(rps0, rps1)
binary_cost = (dist < DISPLACEMENT) * (abs(massdiff) < MASSTHRES)
idx1, idx0 = find_one_to_one_assign(binary_cost)
for i0, i1 in zip(idx0, idx1):
labels[labels1 == rps1[i1].label] = rps0[i0].label
labels0[labels0 == rps0[i0].label] = -rps0[i0].label
return labels0, labels
def _wd(labels0, labels, img0, img1):
labels1 = -labels.copy()
rps0 = regionprops(labels0, img0)
from subdetect_operation import watershed_divide # DO NOT MOVE IT
from utils.track_utils import _find_match
untracked_labels = labels1.copy()
untracked_labels[untracked_labels < 0] = 0
wshed_labels = watershed_divide(untracked_labels,
regmax=REGMAX,
min_size=MIN_SIZE)
wshed_labels = label(wshed_labels)
store = regionprops(wshed_labels, img1)
good_cells = _find_match(rps0, store, DISPLACEMENT, MASSTHRES)
for gc in good_cells: # needed to reuse _update_labels_neck_cut
gccrds = gc.coords[0]
gc.raw_label = labels1[gccrds[0], gccrds[1]]
labels0, labels = _update_labels_neck_cut(labels0, labels1, good_cells)
labels0, labels = nn_closer(img0, img1, labels0, -labels, DISPLACEMENT,
MASSTHRES)
return labels0, labels, good_cells
def track_neck_cut(img0,
img1,
labels0,
labels1,
DISPLACEMENT=10,
MASSTHRES=0.2,
EDGELEN=5,
THRES_ANGLE=180,
WSLIMIT=False,
SMALL_RAD=3,
CANDS_LIMIT=300):
"""
Adaptive segmentation by using tracking informaiton.
Separate two objects by making a cut at the deflection. For each points on the outline,
it will make a triangle separated by EDGELEN and calculates the angle facing inside of concave.
The majority of cells need to be tracked before the this method to calculate LARGE_RAD and SMALL_RAD.
EDGELEN (int): A length of edges of triangle on the nuclear perimeter.
THRES_ANGLE (int): Define the neck points if a triangle has more than this angle.
STEPLIM (int): points of neck needs to be separated by at least STEPLIM in parimeters.
WSLIMIT (bool): Limit search points to ones overlapped with watershed transformed images. Set it True if calculation is slow.
SMALL_RAD (int or None): The smallest radius of candidate objects. If you have many cells, set it to None will infer the radius from previous frame.
CANDS_LIMIT(int): use lower if slow. limit a number of searches.
"""
labels0, labels = nn_closer(img0, img1, labels0, labels1, DISPLACEMENT,
MASSTHRES)
labels1 = -labels.copy()
if SMALL_RAD is None and not hasattr(holder, 'SMALL_RAD'):
tracked_area = [i.area for i in regionprops(labels)]
holder.SMALL_RAD = np.sqrt(np.percentile(tracked_area, 5) / np.pi)
elif SMALL_RAD is not None:
holder.SMALL_RAD = SMALL_RAD
SMALL_RAD = holder.SMALL_RAD
rps0 = regionprops(labels0, img0)
unique_labels = np.unique(labels1)
if WSLIMIT:
wlines = wshed_raw(labels1 > 0, img1)
else:
wlines = np.ones(labels1.shape, np.bool)
store = []
coords_store = []
for label_id in unique_labels:
if label_id == 0:
continue
cc = CellCutter(labels1 == label_id,
img1,
wlines,
small_rad=SMALL_RAD,
EDGELEN=EDGELEN,
THRES=THRES_ANGLE,
CANDS_LIMIT=CANDS_LIMIT)
cc.prepare_coords_set()
candidates = cc.search_cut_candidates(cc.bw.copy(),
cc.coords_set[:CANDS_LIMIT])
for c in candidates:
c.raw_label = label_id
store.append(candidates)
coords_store.append(cc.coords_set)
coords_store = [i for i in coords_store if i]
# Attempt a first cut.
good_cells = _find_best_neck_cut(rps0, store, DISPLACEMENT, MASSTHRES)
labels0, labels = _update_labels_neck_cut(labels0, labels1, good_cells)
labels0, labels = nn_closer(img0, img1, labels0, -labels, DISPLACEMENT,
MASSTHRES)
# iteration from here.
while good_cells:
rps0 = regionprops(labels0, img0)
labels1 = -labels.copy()
rps0 = regionprops(labels0, img0)
unique_labels = np.unique(labels1)
store = []
for label_id in unique_labels:
if label_id == 0:
continue
bw = labels1 == label_id
coords_set = [
i for i in coords_store if bw[i[0][0][0], i[0][0][1]]
]
if not coords_set:
continue
coords_set = coords_set[0]
candidates = cc.search_cut_candidates(bw, coords_set)
for c in candidates:
c.raw_label = label_id
store.append(candidates)
coords_store.append(coords_set)
good_cells = _find_best_neck_cut(rps0, store, DISPLACEMENT, MASSTHRES)
labels0, labels = _update_labels_neck_cut(labels0, labels1, good_cells)
labels0, labels = nn_closer(img0, img1, labels0, -labels, DISPLACEMENT,
MASSTHRES)
return labels0, labels
