def process(self,
input=None,
mode=None,
_DEBUG=False,
_VISUAL_DEBUG=False,
output_folder=tempfile.gettempdir(),
output_name='handCorrection.tif',
threshold=None,
filter=None,
correction_factor=2,
**kwargs):
if input is None:
logger.error('no input image --> nothing to do')
return
# TODO test it with several images just to see if that works
if isinstance(mode, str) and 'first' in mode:
# return first channel only # shall I had a channel axis to it to avoid issues
out = input[..., 0]
# I do this to keep the ...hwc format...
return out[..., np.newaxis]
img_orig = input
if not img_orig.has_c() or img_orig.shape[-1] != 7:
# TODO in fact could do the fast mode still on a single image --> may be useful
logger.error(
'image must have 7 channels to be used for post process')
return img_orig
if _DEBUG:
Img(img_orig, dimensions='hwc').save(
os.path.join(output_folder, 'raw_input.tif'))
bckup_img_wshed = img_orig[..., 0].copy()
if mode is not None and isinstance(mode, str):
if 'ast' in mode:
logger.debug('fast mode')
img_orig[..., 0] += img_orig[..., 1]
img_orig[..., 0] += img_orig[..., 2]
img_orig = img_orig[..., 0] / 3
img_orig = np.reshape(img_orig, (*img_orig.shape, 1))
else:
logger.debug('normal mode')
else:
logger.debug('normal mode')
differing_bonds = np.zeros_like(img_orig)
img_orig[..., 0] = segment_cells(img_orig[..., 0],
min_threshold=0.02,
min_unconnected_object_size=3)
if img_orig.shape[-1] >= 5:
img_orig[..., 1] = segment_cells(img_orig[..., 1],
min_threshold=0.06,
min_unconnected_object_size=6)
img_orig[..., 2] = segment_cells(img_orig[..., 2],
min_threshold=0.15,
min_unconnected_object_size=12)
img_orig[..., 3] = Img.invert(img_orig[..., 3])
img_orig[..., 3] = segment_cells(img_orig[..., 3],
min_threshold=0.06,
min_unconnected_object_size=6)
img_orig[..., 4] = Img.invert(img_orig[..., 4])
img_orig[..., 4] = segment_cells(img_orig[..., 4],
min_threshold=0.15,
min_unconnected_object_size=12)
if img_orig.shape[-1] == 7:
img_orig[..., 5] = self.binarise(img_orig[..., 5], threshold=0.15)
img_orig[..., 6] = Img.invert(img_orig[..., 6])
img_orig[..., 6] = self.binarise(img_orig[..., 6], threshold=0.1)
if _DEBUG:
Img(img_orig, dimensions='hwc').save(
os.path.join(output_folder, 'thresholded_masks.tif'))
# get watershed mask for all images
for i in range(img_orig.shape[-1]):
if i < 5:
final_seeds = label(Img.invert(img_orig[..., i]),
connectivity=1,
background=0)
else:
final_seeds = label(img_orig[..., i],
connectivity=None,
background=0)
final_wshed = watershed(bckup_img_wshed,
markers=final_seeds,
watershed_line=True)
final_wshed[final_wshed != 0] = 1
final_wshed[final_wshed == 0] = 255
final_wshed[final_wshed == 1] = 0
differing_bonds[..., i] = final_wshed
del final_seeds
del final_wshed
if _DEBUG:
print(os.path.join(output_folder, 'differences.tif'))
Img(differing_bonds, dimensions='hwc').save(
os.path.join(output_folder, 'differences.tif'))
Img(bckup_img_wshed, dimensions='hw').save(
os.path.join(output_folder, 'orig_img.tif'))
avg = np.mean(differing_bonds, axis=-1)
avg = avg / avg.max()
if _DEBUG:
Img(avg, dimensions='hw').save(
os.path.join(output_folder, output_name + str('avg.tif')))
if threshold is None:
threshold = self.autothreshold(avg)
logger.debug('threshold used for producing the final mask=' +
str(threshold))
final_mask = avg.copy()
final_mask = self.binarise(final_mask, threshold=threshold)
if _DEBUG:
Img(final_mask, dimensions='hw').save(
os.path.join(output_folder, 'binarized.tif'))
# close wshed mask to fill super tiny holes
s = ndimage.generate_binary_structure(2, 1)
final_mask = ndimage.grey_dilation(final_mask, footprint=s)
# remove super tiny artificial cells (very small value cause already dilated)
mask = label(Img.invert(final_mask), connectivity=1, background=0)
for region in regionprops(mask):
if region.area < 5:
for coordinates in region.coords:
final_mask[coordinates[0], coordinates[1]] = 255
del mask
final_mask = label(Img.invert(final_mask),
connectivity=1,
background=0)
final_mask = watershed(bckup_img_wshed,
markers=final_mask,
watershed_line=True)
final_mask[final_mask != 0] = 1
final_mask[final_mask == 0] = 255
final_mask[final_mask == 1] = 0
if filter is None or filter == 0:
return final_mask.astype(np.uint8)
else:
logger.debug('Further filtering image')
return FilterMask(bckup_img_wshed,
final_mask,
filter=filter,
correction_factor=correction_factor)
def FilterMask(img_orig, final_wshed, filter='local median', correction_factor=2,
_DEBUG=False,
_VISUAL_DEBUG=False,
**kwargs):
labs = label(Img.invert(final_wshed.astype(np.uint8)), connectivity=1, background=0)
start = timer()
output_folder = '/home/aigouy/Bureau/trash/test_new_seeds_seg_stuff/'
if filter is None or filter == 0:
return final_wshed.astype(np.uint8)
else:
if isinstance(filter, int):
filter_by_size = filter
else:
filter_by_size = None
avg_area = 0
count = 0
if _DEBUG:
Img(final_wshed, dimensions='hw').save(os.path.join(output_folder, 'extras', 'test_size_cells.tif'))
final_seeds = Img.invert(final_wshed)
final_seeds = label(final_seeds, connectivity=1, background=0)
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
removed_seeds = []
keep_seeds = []
labels_n_bbox = {}
labels_n_area = {}
border_cells = []
ids_n_local_median = {}
if isinstance(filter, str) and 'local' in filter:
rps = regionprops(final_seeds)
for region in rps:
labels_n_bbox[region.label] = region.bbox
labels_n_area[region.label] = region.area
if (region.bbox[0] <= 3 or region.bbox[1] <= 3 or region.bbox[2] >= final_seeds.shape[-2] - 5 or
region.bbox[
3] >= \
final_seeds.shape[-1] - 5):
border_cells.append(region.label)
_, tiles = Img.get_2D_tiles_with_overlap(final_seeds, overlap=64, dimension_h=-2, dimension_w=-1)
for r in tiles:
for tile in r:
rps2 = regionprops(tile)
for region in rps2:
if region.label in border_cells:
continue
if (region.bbox[0] <= 3 or region.bbox[1] <= 3 or region.bbox[2] >= final_seeds.shape[
-2] - 5 or
region.bbox[
3] >= \
final_seeds.shape[-1] - 5):
continue
area_of_neighboring_cells = []
for region2 in rps2:
if region2.label == region.label:
continue
# find all cells with
if rect_distance(region.bbox, region2.bbox) <= 1:
area_of_neighboring_cells.append(labels_n_area[region2.label])
if area_of_neighboring_cells:
median = statistics.median_low(area_of_neighboring_cells)
ids_n_local_median[
region.label] = median / correction_factor
if region.area <= median / correction_factor:
removed_seeds.append(region.label)
else:
keep_seeds.append(region.label)
removed_seeds = [x for x in removed_seeds if x not in keep_seeds]
# TODO offer the things below as an option --> prevent removal of sure seeds or something like that
else:
areas = []
for region in regionprops(final_seeds):
if (region.bbox[0] <= 3 or region.bbox[1] <= 3 or region.bbox[2] >= final_seeds.shape[-2] - 5 or
region.bbox[3] >= final_seeds.shape[-1] - 5):
continue
avg_area += region.area
count += 1
areas.append(region.area)
avg_area /= count
median = statistics.median_low(areas)
if isinstance(filter, int):
filter_by_size = filter
elif 'avg' in filter:
filter_by_size = avg_area / correction_factor
elif 'median' in filter:
filter_by_size = median / correction_factor
# TODO maybe use stdev or alike to see if cell should really be removed
if _DEBUG:
print('filter cells below=', filter_by_size, 'avg cell area=', avg_area, 'median=',
median) # , 'median', median
if filter_by_size is not None and filter_by_size != 0:
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
for region in regionprops(final_seeds):
labels_n_bbox[region.label] = region.bbox
labels_n_area[region.label] = region.area
if region.area < filter_by_size:
if (region.bbox[0] <= 2 or region.bbox[1] <= 2 or region.bbox[2] >= labs.shape[
-2] - 3 or
region.bbox[
3] >= \
labs.shape[
-1] - 3):
continue
removed_seeds.append(region.label)
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
for region in regionprops(final_seeds):
if region.label in removed_seeds:
for coordinates in region.coords:
final_seeds[coordinates[0], coordinates[1]] = 0
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
final_wshed = watershed(img_orig, markers=final_seeds, watershed_line=True)
final_wshed[final_wshed != 0] = 1 # remove all seeds
final_wshed[final_wshed == 0] = 255 # set wshed values to 255
final_wshed[final_wshed == 1] = 0 # set all other cell content to
if _VISUAL_DEBUG:
plt.imshow(final_wshed)
plt.show()
duration = timer() - start
if _DEBUG:
print('final duration wshed in secs', duration)
return final_wshed.astype(np.uint8)
def process(self, input=None, output_folder=None, progress_callback=None, filter=None,
correction_factor=2,
cutoff_cell_fusion=None,
restore_safe_cells=False,
_DEBUG=False,
_VISUAL_DEBUG=False, **kwargs):
start = timer()
# filename0 = path
# filename0_without_path = os.path.basename(filename0)
# filename0_without_ext = os.path.splitext(filename0_without_path)[0]
# parent_dir_of_filename0 = os.path.dirname(filename0)
# TA_output_filename = os.path.join(parent_dir_of_filename0, filename0_without_ext,
# 'handCorrection.tif') # TODO allow custom names here to allow ensemble methods
# non_TA_final_output_name = os.path.join(output_folder, filename0_without_ext + '.tif')
#
# filename_to_use_to_save = non_TA_final_output_name
# if TA_mode:
# filename_to_use_to_save = TA_output_filename
#
# if TA_mode:
# # try also to change path input name
# if os.path.exists(
# os.path.join(parent_dir_of_filename0, filename0_without_ext, 'raw_epyseg_output.tif')):
# path = os.path.join(parent_dir_of_filename0, filename0_without_ext, 'raw_epyseg_output.tif')
# img_orig = Img(path)
# print('analyzing', path, self.stop_now)
# try:
# if self.progress_callback is not None:
# self.progress_callback.emit((iii / len(list_of_files)) * 100)
# else:
# logger.info(str((iii / len(list_of_files)) * 100) + '%')
# except:
# traceback.print_exc()
# pass
# DO A DILATION OF SEEDS THEN AN EROSION TO JOIN CLOSE BY SEEDS
img_orig = input
img_has_seeds = True
# mask with several channels
if img_orig.has_c():
if restore_safe_cells:
img_seg = img_orig[..., 0].copy()
seeds_1 = img_orig[..., img_orig.shape[-1] - 1]
seeds_1 = Img.invert(seeds_1)
# seeds_1[seeds_1 >= 0.5] = 255
# seeds_1[seeds_1 < 0.5] = 0
seeds_1[seeds_1 >= 0.2] = 255 # TODO maybe be more stringent here
seeds_1[seeds_1 < 0.2] = 0
s = ndimage.generate_binary_structure(2, 1)
seeds_1 = ndimage.grey_dilation(seeds_1, footprint=s)
seeds_1 = ndimage.grey_dilation(seeds_1, footprint=s)
seeds_1 = ndimage.grey_dilation(seeds_1, footprint=s)
seeds_1 = ndimage.grey_erosion(seeds_1, footprint=s)
seeds_1 = ndimage.grey_erosion(seeds_1, footprint=s)
# seeds_1 = ndimage.grey_erosion(seeds_1, footprint=s)
# seeds_1 = ndimage.grey_erosion(seeds_1, footprint=s)
# for debug
if _DEBUG:
Img(seeds_1, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'wshed_seeds.tif')) # not bad
lab_seeds = label(seeds_1.astype(np.uint8), connectivity=2, background=0)
#
for region in regionprops(lab_seeds):
if region.area < 10:
for coordinates in region.coords:
lab_seeds[coordinates[0], coordinates[1]] = 0
if _DEBUG:
Img(seeds_1, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'wshed_seeds_deblobed.tif'))
img_orig[..., 3] = Img.invert(img_orig[..., 3])
img_orig[..., 4] = Img.invert(img_orig[..., 4])
# seems to work --> now need to do the projection
for c in range(1, img_orig.shape[-1] - 2):
img_orig[..., 0] += img_orig[..., 1]
img_orig[..., 0] /= img_orig.shape[-1] - 2
img_orig = img_orig[..., 0]
else:
# mask with single channel
img_has_seeds = False
if restore_safe_cells:
img_seg = img_orig.copy()
if restore_safe_cells:
if _DEBUG:
print(os.path.join(output_folder, 'extras', 'img_seg.tif'))
Img(img_seg, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'img_seg.tif'))
# for debug
if _DEBUG:
Img(img_orig, dimensions='hw').save(os.path.join(output_folder, 'extras', 'avg.tif'))
img_saturated = img_orig.copy()
if img_has_seeds:
img_saturated[img_saturated >= 0.5] = 255
img_saturated[img_saturated < 0.5] = 0
if restore_safe_cells:
# TODO maybe do a safe image
img_seg[img_seg >= 0.3] = 255
img_seg[img_seg < 0.3] = 0
secure_mask = img_seg
else:
img_saturated[img_saturated >= 0.3] = 255
img_saturated[img_saturated < 0.3] = 0
if restore_safe_cells:
img_seg[img_seg >= 0.95] = 255
img_seg[img_seg < 0.95] = 0
secure_mask = img_seg
# convert it to seeds and make sure they are all present in there
# if pixel is not labeled then read it
if restore_safe_cells:
labels_n_area_rescue_seeds = {}
rescue_seeds = label(Img.invert(secure_mask), connectivity=1, background=0)
for region in regionprops(rescue_seeds):
labels_n_area_rescue_seeds[region.label] = region.area
if _DEBUG:
Img(secure_mask, dimensions='hw').save(os.path.join(output_folder, 'extras', 'secure_mask.tif'))
# loop over those seeds to rescue
# for debug
if _DEBUG:
Img(img_saturated, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'handCorrection.tif'))
deblob = True
if deblob:
image_thresh = label(img_saturated, connectivity=2, background=0)
# for debug
if _DEBUG:
Img(image_thresh, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'before_deblobed.tif'))
# deblob
min_size = 200
for region in regionprops(image_thresh):
# take regions with large enough areas
if region.area < min_size:
for coordinates in region.coords:
image_thresh[coordinates[0], coordinates[1]] = 0
image_thresh[image_thresh > 0] = 255
img_saturated = image_thresh
# for debug
if _DEBUG:
Img(img_saturated, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'deblobed.tif'))
del image_thresh
# for debug
if _DEBUG:
Img(img_saturated, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'deblobed_out.tif'))
extra_dilations = True
if extra_dilations:
# do a dilation of 2 to close bonds
s = ndimage.generate_binary_structure(2, 1)
dilated = ndimage.grey_dilation(img_saturated, footprint=s)
dilated = ndimage.grey_dilation(dilated, footprint=s)
# Img(dilated, dimensions='hw').save(os.path.join(os.path.splitext(path)[0], 'filled_one_px_holes.tif'))
# other_seeds = label(invert(np.grey_dilation(dilated, footprint=s).astype(np.uint8)), connectivity=1, background=0)
labs = label(Img.invert(img_saturated.astype(np.uint8)), connectivity=1, background=0)
for region in regionprops(labs):
seeds = []
# exclude tiny cells form dilation because they may end up completely closed
if region.area >= 10 and region.area < 350:
for coordinates in region.coords:
dilated[coordinates[0], coordinates[1]] = 0
continue
else:
# pb when big cells around cause connections are not done
# preserve cells at edges because they have to e naturally smaller because they are cut
# put a size criterion too
if region.area < 100 and (
region.bbox[0] <= 1 or region.bbox[1] <= 1 or region.bbox[2] >= labs.shape[-2] - 2 or
region.bbox[
3] >= \
labs.shape[-1] - 2):
# edge cell detected --> removing dilation
for coordinates in region.coords:
dilated[coordinates[0], coordinates[1]] = 0
continue
img_saturated = dilated
# for debug
if _DEBUG:
Img(img_saturated, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'dilated_further.tif'))
del dilated
list_of_cells_to_dilate = []
labs = label(Img.invert(img_saturated.astype(np.uint8)), connectivity=1, background=0)
# c'est cette correction qui fixe bcp de choses mais recree aussi des choses qui n'existent pas... --> voir à quoi sont dus ces lignes blobs
# faudrait redeblober
if img_has_seeds:
for region in regionprops(labs, intensity_image=img_orig):
seeds = []
if not extra_dilations and region.area < 10:
continue
# if small and no associated seeds --> remove it ??? maybe or not
for coordinates in region.coords:
id = lab_seeds[coordinates[0], coordinates[1]]
if id != 0:
seeds.append(id)
seeds = set(seeds)
if len(seeds) >= 2:
# we may have found an undersegmented cell --> try segment it better
list_of_cells_to_dilate.append(region.label)
if len(list_of_cells_to_dilate) != 0:
props = regionprops(labs, intensity_image=img_orig)
for run in range(10):
something_changed = False # early stop
for region in props:
if region.label not in list_of_cells_to_dilate:
continue
# TODO recheck those values and wether it makes sense
threshold_values = [80 / 255, 60 / 255, 40 / 255, 30 / 255,
20 / 255,
10 / 255] # 160 / 255, 140 / 255, 120 / 255, 100 / 255, 1 / 255 , 2 / 255, , 5 / 255
try:
for threshold in threshold_values:
mask = region.image.copy()
image = region.image.copy()
image[region.intensity_image > threshold] = True
image[region.intensity_image <= threshold] = False
final = Img.invert(image.astype(np.uint8))
final[final < 255] = 0
final[mask == False] = 0
new_seeds = label(final, connectivity=1, background=0)
props2 = regionprops(new_seeds)
if len(props2) > 1: # cell was resplitted into smaller
for r in props2:
if r.area < 20:
raise Exception
region.image[mask == False] = False
region.image[mask == True] = True
region.image[new_seeds > 0] = False
something_changed = True
for coordinates in region.coords:
img_saturated[coordinates[0], coordinates[1]] = 255
region.image[mask == False] = False
region.image[mask == True] = True
del final
del new_seeds
except:
traceback.print_exc()
pass
if not something_changed:
# print('no more changes anymore --> quitting')
break
# for debug
if _DEBUG:
Img(img_saturated, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'saturated_mask4.tif'))
final_seeds = label(Img.invert(img_saturated), connectivity=1,
background=0) # keep like that otherwise creates tiny cells with erroneous wshed
# for debug
if _DEBUG:
Img(final_seeds, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'final_seeds_before.tif'))
final_seeds = label(Img.invert(img_saturated), connectivity=2, background=0) # is that needed ???
# for debug
if _DEBUG:
Img(final_seeds, dimensions='hw').save(
os.path.join(output_folder, 'extras', 'final_seeds_before2.tif'))
final_seeds[img_saturated == 255] = 0
final_wshed = watershed(img_orig, markers=final_seeds,
watershed_line=True)
final_wshed[final_wshed != 0] = 1 # remove all seeds
final_wshed[final_wshed == 0] = 255 # set wshed values to 255
final_wshed[final_wshed == 1] = 0 # set all other cell content to
# filename0 = os.path.basename(path)
# parent_path = os.path.dirname(os.path.dirname(path))
if filter is None or filter == 0:
# TODO maybe offer the choice between saving wshed on predict or on orig
# Img(final_wshed, dimensions='hw').save(os.path.join(output_folder, os.path.splitext(filename0)[
# 0]) + '.tif') # need put original name here TODO put image default name here
# print('saving', filename_to_use_to_save)
# Img(final_wshed.astype(np.uint8), dimensions='hw').save(filename_to_use_to_save)
return final_wshed.astype(np.uint8)
else:
if isinstance(filter, int):
filter_by_size = filter
else:
filter_by_size = None
avg_area = 0
count = 0
if _DEBUG:
Img(final_wshed, dimensions='hw').save(os.path.join(output_folder, 'extras', 'test_size_cells.tif'))
final_seeds = Img.invert(final_wshed)
final_seeds = label(final_seeds, connectivity=1, background=0)
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
removed_seeds = []
keep_seeds = []
labels_n_bbox = {}
labels_n_area = {}
border_cells = []
ids_n_local_median = {}
correspondance_between_cur_seeds_and_safe_ones = {}
if isinstance(filter, str) and 'local' in filter:
rps = regionprops(final_seeds)
for region in rps:
labels_n_bbox[region.label] = region.bbox
labels_n_area[region.label] = region.area
if (region.bbox[0] <= 3 or region.bbox[1] <= 3 or region.bbox[2] >= final_seeds.shape[-2] - 5 or
region.bbox[
3] >= \
final_seeds.shape[-1] - 5):
border_cells.append(region.label)
if restore_safe_cells:
for coordinates in region.coords:
if rescue_seeds[coordinates[0], coordinates[1]] != 0: # do r
correspondance_between_cur_seeds_and_safe_ones[region.label] = rescue_seeds[
coordinates[0], coordinates[1]]
break
break
_, tiles = Img.get_2D_tiles_with_overlap(final_seeds, overlap=64, dimension_h=-2, dimension_w=-1)
for r in tiles:
for tile in r:
rps2 = regionprops(tile)
for region in rps2:
if self.stop_now:
return
if region.label in border_cells:
continue
if (region.bbox[0] <= 3 or region.bbox[1] <= 3 or region.bbox[2] >= final_seeds.shape[
-2] - 5 or
region.bbox[
3] >= \
final_seeds.shape[-1] - 5):
continue
area_of_neighboring_cells = []
for region2 in rps2:
if region2.label == region.label:
continue
# find all cells with
if self.rect_distance(region.bbox, region2.bbox) <= 1:
area_of_neighboring_cells.append(labels_n_area[region2.label])
if area_of_neighboring_cells:
median = statistics.median_low(area_of_neighboring_cells)
ids_n_local_median[
region.label] = median / correction_factor
if region.area <= median / correction_factor:
removed_seeds.append(region.label)
else:
keep_seeds.append(region.label)
removed_seeds = [x for x in removed_seeds if x not in keep_seeds]
# TODO offer the things below as an option --> prevent removal of sure seeds or something like that
if restore_safe_cells:
removed_seeds_to_restore = []
for region in regionprops(final_seeds):
if region.label in removed_seeds:
first = True
for coordinates in region.coords:
if first and rescue_seeds[coordinates[0], coordinates[1]] != 0:
percent_diff = min(labels_n_area[region.label], labels_n_area_rescue_seeds[
rescue_seeds[coordinates[0], coordinates[1]]]) / max(
labels_n_area[region.label], labels_n_area_rescue_seeds[
rescue_seeds[coordinates[0], coordinates[1]]])
if (percent_diff >= 0.7 and percent_diff < 1.0) or (
labels_n_area[region.label] <= 200 and (
percent_diff >= 0.3 and percent_diff < 1.0)):
if _DEBUG:
print('0 finally not removing seed, safe seed', region.label,
percent_diff,
labels_n_area[region.label],
labels_n_area_rescue_seeds[
rescue_seeds[coordinates[0], coordinates[1]]],
labels_n_area[region.label] / labels_n_area_rescue_seeds[
rescue_seeds[coordinates[0], coordinates[1]]],
region.centroid)
removed_seeds_to_restore.append(region.label)
break
break
removed_seeds = [x for x in removed_seeds if x not in removed_seeds_to_restore]
else:
areas = []
for region in regionprops(final_seeds):
if (region.bbox[0] <= 3 or region.bbox[1] <= 3 or region.bbox[2] >= final_seeds.shape[-2] - 5 or
region.bbox[3] >= final_seeds.shape[-1] - 5):
continue
avg_area += region.area
count += 1
areas.append(region.area)
avg_area /= count
median = statistics.median_low(areas)
if isinstance(filter, int):
filter_by_size = filter
elif 'avg' in filter:
filter_by_size = avg_area / correction_factor
elif 'median' in filter:
filter_by_size = median / correction_factor
# TODO maybe use stdev or alike to see if cell should really be removed
if _DEBUG:
print('filter cells below=', filter_by_size, 'avg cell area=', avg_area, 'median=',
median) # , 'median', median
if filter_by_size is not None and filter_by_size != 0:
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
for region in regionprops(final_seeds):
labels_n_bbox[region.label] = region.bbox
labels_n_area[region.label] = region.area
if region.area < filter_by_size:
if (region.bbox[0] <= 2 or region.bbox[1] <= 2 or region.bbox[2] >= labs.shape[
-2] - 3 or
region.bbox[
3] >= \
labs.shape[
-1] - 3):
continue
removed_seeds.append(region.label)
if cutoff_cell_fusion is not None and cutoff_cell_fusion > 1:
cells_to_fuse = []
for idx, removed_seed in enumerate(removed_seeds):
current_cells_to_fuse = set()
closest_pair = None
smallest_distance = None
for idx2 in range(idx + 1, len(removed_seeds)):
removed_seed2 = removed_seeds[idx2]
if closest_pair is None:
if self.rect_distance(labels_n_bbox[removed_seed], labels_n_bbox[removed_seed2]) <= 1:
closest_pair = removed_seed2
smallest_distance = self.rect_distance(labels_n_bbox[removed_seed],
labels_n_bbox[removed_seed2])
elif self.rect_distance(labels_n_bbox[removed_seed],
labels_n_bbox[removed_seed2]) <= smallest_distance:
closest_pair = removed_seed2
smallest_distance = self.rect_distance(labels_n_bbox[removed_seed],
labels_n_bbox[removed_seed2])
if self.rect_distance(labels_n_bbox[removed_seed], labels_n_bbox[removed_seed2]) <= 1:
current_cells_to_fuse.add(removed_seed)
current_cells_to_fuse.add(removed_seed2)
if current_cells_to_fuse:
cells_to_fuse.append(current_cells_to_fuse)
cells_to_fuse = [frozenset(i) for i in cells_to_fuse]
cells_to_fuse = list(dict.fromkeys(cells_to_fuse))
cells_to_keep = []
if cutoff_cell_fusion is not None and cutoff_cell_fusion > 0:
superfuse = []
copy_of_cells_to_fuse = cells_to_fuse.copy()
for idx, fuse in enumerate(copy_of_cells_to_fuse):
current_fusion = set(fuse.copy())
changed = True
while changed:
changed = False
for idx2 in range(len(copy_of_cells_to_fuse) - 1, idx, -1):
fuse2 = copy_of_cells_to_fuse[idx2]
if idx2 == idx:
continue
if fuse2.intersection(current_fusion):
current_fusion.update(fuse2)
del copy_of_cells_to_fuse[idx2]
changed = True
superfuse.append(current_fusion)
for sf in superfuse:
if len(sf) > cutoff_cell_fusion:
for val in sf:
cells_to_keep.append(val)
seeds_to_fuse = []
cells_to_fuse = sorted(cells_to_fuse, key=len)
for fuse in cells_to_fuse:
cumulative_area = 0
for _id in fuse:
if _id in cells_to_keep:
if _id in removed_seeds:
removed_seeds.remove(_id)
continue
cumulative_area += labels_n_area[_id]
if filter_by_size is not None:
if cumulative_area >= filter_by_size: #: #1200: #filter_by_size: # need hack this to get local area
seeds_to_fuse.append(fuse)
for _id in fuse:
if _id in removed_seeds:
removed_seeds.remove(_id)
else:
if cumulative_area >= ids_n_local_median[_id]:
seeds_to_fuse.append(fuse)
for _id in fuse:
if _id in removed_seeds:
removed_seeds.remove(_id)
# need recolor all the seeds in there with the new seed stuff
for fuse in seeds_to_fuse:
for _id in fuse:
break
for region in regionprops(final_seeds):
if region.label in fuse:
for coordinates in region.coords:
final_seeds[coordinates[0], coordinates[1]] = _id
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
for region in regionprops(final_seeds):
if region.label in removed_seeds:
for coordinates in region.coords:
final_seeds[coordinates[0], coordinates[1]] = 0
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
if _VISUAL_DEBUG:
plt.imshow(final_seeds)
plt.show()
final_wshed = watershed(img_orig, markers=final_seeds, watershed_line=True)
final_wshed[final_wshed != 0] = 1 # remove all seeds
final_wshed[final_wshed == 0] = 255 # set wshed values to 255
final_wshed[final_wshed == 1] = 0 # set all other cell content to
if _VISUAL_DEBUG:
plt.imshow(final_wshed)
plt.show()
# print('saving', filename_to_use_to_save)
# Img(final_wshed.astype(np.uint8), dimensions='hw').save(filename_to_use_to_save)
duration = timer() - start
if _DEBUG:
print('final duration wshed in secs', duration)
return final_wshed.astype(np.uint8) # is indeed a 2D image
def get_optimized_mask2(img,
sauvola_mask=None,
use_quick_shift=False,
__VISUAL_DEBUG=False,
__DEBUG=False,
score_before_adding=False,
return_seeds=False):
final_image = None
if use_quick_shift:
rotations = [0, 2, 3]
kernels = [1, 1.33, 1.66, 2]
for kern in kernels:
for rot in rotations:
segments_quick = getQuickseg(img,
nb_of_90_rotation=rot,
kernel_size=kern)
segments_quick = segments_quick.astype(np.uint8)
if final_image is None:
final_image = segments_quick
else:
final_image = final_image + segments_quick
if __VISUAL_DEBUG:
plt.imshow(final_image)
plt.title("avg")
plt.show()
final_image[final_image < final_image.max()] = 0
final_image[final_image >= final_image.max()] = 1
if sauvola_mask is None:
from epyseg.postprocess.edmshed import sauvola
t = sauvola(img, min_threshold=0.1, window_size=25)
img[img >= t] = 1
img[img < t] = 0
else:
img = sauvola_mask
img[img > 0] = 1
if __DEBUG:
Img(img, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/sauvola_mask.tif')
raw_sauvola = img.copy()
if use_quick_shift:
final_image[raw_sauvola != 0] = 1
if __VISUAL_DEBUG:
plt.imshow(final_image)
plt.show()
if __DEBUG:
Img(final_image.astype(np.uint8) * 255, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/corrected_stuff.tif')
final_image = ~remove_small_objects(
~final_image.astype(np.bool), min_size=5, connectivity=1)
final_image = skeletonize(final_image)
if __DEBUG:
Img(final_image.astype(np.uint8), dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/skel_quick.tif')
if __VISUAL_DEBUG:
plt.imshow(final_image)
plt.title("binary")
plt.show()
vertices_quick, cut_bonds_quick = split_into_vertices_and_bonds(
final_image)
if __DEBUG:
Img(vertices_quick.astype(np.uint8), dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/vertices_quick.tif')
Img(cut_bonds_quick.astype(np.uint8), dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/cut_bonds_quick.tif')
img = skeletonize(img)
if __DEBUG:
Img(img, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/skel_sauvola_mask.tif')
img = remove_small_objects(img, min_size=6, connectivity=2)
if __DEBUG:
Img(img, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/skel_sauvola_mask_deblobed.tif')
if __VISUAL_DEBUG:
plt.imshow(img)
plt.show()
image = img.copy()
image = image.astype(np.uint8) * 255
image = Img.invert(image)
distance = distance_transform_edt(image)
if __VISUAL_DEBUG:
plt.imshow(distance)
plt.show()
local_maxi = peak_local_max(distance,
indices=False,
footprint=np.ones((8, 8)),
labels=image)
distance = -distance
markers = ndimage.label(local_maxi,
structure=generate_binary_structure(2, 2))[0]
if __VISUAL_DEBUG:
plt.imshow(markers)
plt.show()
if __DEBUG:
Img(markers, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/markers_0.tif')
labels = watershed(distance, markers,
watershed_line=True) # --> maybe implement that too
labels[labels != 0] = 1
labels[labels == 0] = 255
labels[labels == 1] = 0
labels[labels == 255] = 1
if __VISUAL_DEBUG:
plt.imshow(labels)
plt.title('raw wshed')
plt.show()
if __DEBUG:
Img(labels.astype(np.uint8), dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/wshed_edm.tif')
vertices_edm, cut_bonds_edm = split_into_vertices_and_bonds(labels)
if __DEBUG:
Img(vertices_edm.astype(np.uint8), dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/vertices_edm.tif')
Img(cut_bonds_edm.astype(np.uint8), dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/cut_bonds_edm.tif')
if use_quick_shift:
if __DEBUG:
Img(img, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/skel_sauvola_mask_deblobed.tif'
)
unconnected = detect_unconnected_bonds(img)
if __DEBUG:
Img(unconnected, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/unconnected.tif')
labels_quick = label(cut_bonds_quick, connectivity=2, background=0)
labels_quick_vertices = label(vertices_quick,
connectivity=2,
background=0)
props_labels_quick = regionprops(labels_quick)
labels_pred = label(unconnected, connectivity=2, background=0)
raw_sauvola = rescue_bonds(labels_pred,
labels_quick,
raw_sauvola,
labels_quick_vertices,
props_labels_quick,
score_before_adding=score_before_adding)
if __DEBUG:
Img(raw_sauvola.astype(np.uint8), dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/corrected_bonds_sauvola.tif')
labels_edm = label(cut_bonds_edm, connectivity=2, background=0)
labels_edm_vertices = label(vertices_edm, connectivity=2, background=0)
props_labels_edm = regionprops(labels_edm)
img = skeletonize(raw_sauvola)
if __DEBUG:
Img(raw_sauvola, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/skel_sauvola_mask2.tif')
unconnected = detect_unconnected_bonds(img)
if __DEBUG:
Img(unconnected, dimensions='hw').save(
'/home/aigouy/Bureau/trash/trash4/unconnected.tif')
labels_pred = label(unconnected, connectivity=2, background=0)
raw_sauvola = rescue_bonds(labels_pred,
labels_edm,
raw_sauvola,
labels_edm_vertices,
props_labels_edm,
score_before_adding=score_before_adding)
raw_sauvola = connect_unconnected(labels_pred, labels_edm, raw_sauvola,
props_labels_edm, labels_edm_vertices)
if return_seeds:
return markers, labels, raw_sauvola
return raw_sauvola