def CalculateSomataStatistics(meta_filename):
data = ReadMetaData(meta_filename)
somata_statistics = {}
# iterate over all blocks
for iz in range(data.StartZ(), data.EndZ()):
for iy in range(data.StartY(), data.EndY()):
for ix in range(data.StartX(), data.EndX()):
print('{} {:04d}z-{:04d}y-{:04d}x'.format(
meta_filename, iz, iy, ix))
# some datasets have no somata (default value)
upsampled_non_zero_voxels = 0
if data.SomataDownsampleRate():
somata = data.ReadSomataBlock(iz, iy, ix)
# get the number of non zero voxels
non_zero_voxels = np.count_nonzero(somata)
# the upsample factor is the number of voxels at full resolution correspond
# to one voxel at the downsampled resolution
upsample_factor = data.SomataDownsampleRate()**3
# the number of voxels masked as full resolution
upsampled_non_zero_voxels = upsample_factor * non_zero_voxels
somata_statistics[(iz, iy, ix)] = upsampled_non_zero_voxels
statistics_directory = '{}/statistics'.format(data.TempDirectory())
if not os.path.exists(statistics_directory):
os.makedirs(statistics_directory, exist_ok=True)
statistics_filename = '{}/somata-statistics.pickle'.format(
statistics_directory)
PickleData(somata_statistics, statistics_filename)
def EvaluateNeuralReconstructionIntegrity(meta_filename):
data = ReadMetaData(meta_filename)
synapses_per_label = {}
# read in all of the synapses from all of the blocks
synapse_directory = data.SynapseDirectory()
for iz in range(data.StartZ(), data.EndZ()):
for iy in range(data.StartY(), data.EndY()):
for ix in range(data.StartX(), data.EndX()):
synapses_filename = '{}/{:04d}z-{:04d}y-{:04d}x.pts'.format(synapse_directory, iz, iy, ix)
# ignore the local coordinates
block_synapses, _ = ReadPtsFile(data, synapses_filename)
# add all synapses for each label in this block to the global synapses
for label in block_synapses.keys():
if not label in synapses_per_label:
synapses_per_label[label] = []
synapses_per_label[label] += block_synapses[label]
# get the output filename
evaluation_directory = data.EvaluationDirectory()
if not os.path.exists(evaluation_directory):
os.makedirs(evaluation_directory, exist_ok=True)
output_filename = '{}/nri-results.txt'.format(evaluation_directory)
fd = open(output_filename, 'w')
# keep track of global statistics
total_true_positives = 0
total_false_positives = 0
total_false_negatives = 0
# for each label, find if synapses correspond to endpoints
for label in range(1, data.NLabels()):
# read the refined skeleton for this synapse
skeleton_directory = '{}/skeletons'.format(data.SkeletonOutputDirectory())
skeleton_filename = '{}/{:016d}.pts'.format(skeleton_directory, label)
# skip over labels not processed
if not os.path.exists(skeleton_filename): continue
# get the synapses only for this one label
synapses = synapses_per_label[label]
# ignore the local coordinates
skeletons, _ = ReadPtsFile(data, skeleton_filename)
skeleton = skeletons[label]
# read in the somata surfaces (points on the surface should not count as endpoints)
somata_directory = '{}/somata_surfaces'.format(data.TempDirectory())
somata_filename = '{}/{:016d}.pts'.format(somata_directory, label)
# path may not exist if soma not found
if os.path.exists(somata_filename):
somata_surfaces, _ = ReadPtsFile(data, somata_filename)
somata_surface = set(somata_surfaces[label])
else:
somata_surface = set()
# get the endpoints in this skeleton for this label
endpoints = FindEndpoints(data, skeleton, somata_surface)
true_positives, false_positives, false_negatives = CalculateNeuralReconstructionIntegrityScore(data, synapses, endpoints)
# if there are no true positives the NRI score is 0
if true_positives == 0:
nri_score = 0
else:
precision = true_positives / float(true_positives + false_positives)
recall = true_positives / float(true_positives + false_negatives)
nri_score = 2 * (precision * recall) / (precision + recall)
print ('Label: {}'.format(label))
print (' True Positives: {:10d}'.format(true_positives))
print (' False Positives: {:10d}'.format(false_positives))
print (' False Negatives: {:10d}'.format(false_negatives))
print (' NRI Score: {:0.4f}'.format(nri_score))
fd.write ('Label: {}\n'.format(label))
fd.write (' True Positives: {:10d}\n'.format(true_positives))
fd.write (' False Positives: {:10d}\n'.format(false_positives))
fd.write (' False Negatives: {:10d}\n'.format(false_negatives))
fd.write (' NRI Score: {:0.4f}\n'.format(nri_score))
# update the global stats
total_true_positives += true_positives
total_false_positives += false_positives
total_false_negatives += false_negatives
precision = total_true_positives / float(total_true_positives + total_false_positives)
recall = total_true_positives / float(total_true_positives + total_false_negatives)
nri_score = 2 * (precision * recall) / (precision + recall)
print ('Total Volume'.format(label))
print (' True Positives: {:10d}'.format(total_true_positives))
print (' False Positives: {:10d}'.format(total_false_positives))
print (' False Negatives: {:10d}'.format(total_false_negatives))
print (' NRI Score: {:0.4f}'.format(nri_score))
fd.write ('Total Volume\n'.format(label))
fd.write (' True Positives: {:10d}\n'.format(total_true_positives))
fd.write (' False Positives: {:10d}\n'.format(total_false_positives))
fd.write (' False Negatives: {:10d}\n'.format(total_false_negatives))
fd.write (' NRI Score: {:0.4f}\n'.format(nri_score))
def EvaluateHoleFilling(meta_filename):
data = ReadMetaData(meta_filename)
# make sure a results folder is specified
assert (not data.EvaluationDirectory() == None)
hole_sizes = {}
neighbor_label_dicts = {}
# read in the hole sizes from each block
for iz in range(data.StartZ(), data.EndZ()):
for iy in range(data.StartY(), data.EndY()):
for ix in range(data.StartX(), data.EndX()):
tmp_block_directory = data.TempBlockDirectory(iz, iy, ix)
# read the saved hole sizes for this block
hole_sizes_filename = '{}/hole-sizes.pickle'.format(tmp_block_directory)
holes_sizes_per_block = ReadPickledData(hole_sizes_filename)
for label in holes_sizes_per_block:
hole_sizes[label] = holes_sizes_per_block[label]
# any value already determined in the local step mush have no neighbors
associated_label_dict = ReadPickledData('{}/associated-label-set-local.pickle'.format(tmp_block_directory))
for label in associated_label_dict:
neighbor_label_dicts[label] = []
# read in the neighbor label dictionary that maps values to its neighbors
tmp_directory = data.TempDirectory()
neighbor_label_filename = '{}/hole-filling-neighbor-label-dict-global.pickle'.format(tmp_directory)
neighbor_label_dict_global = ReadPickledData(neighbor_label_filename)
associated_label_filename = '{}/hole-filling-associated-labels.pickle'.format(tmp_directory)
associated_label_dict = ReadPickledData(associated_label_filename)
# make sure that the keys are identical for hole sizes and associated labels (sanity check)
assert (sorted(hole_sizes.keys()) == sorted(associated_label_dict.keys()))
# make sure no query component in the global dictionary occurs in the local dictionary
for label in neighbor_label_dict_global.keys():
assert (not label in neighbor_label_dicts)
# create a unified neighbor labels dictionary that combines local and global information
neighbor_label_dicts.update(neighbor_label_dict_global)
# make sure that the keys are identical for hole sizes and the neighbor label dicts
assert (sorted(hole_sizes.keys()) == sorted(neighbor_label_dicts.keys()))
# union find data structure to link together holes across blocks
class UnionFindElement:
def __init__(self, label):
self.label = label
self.parent = self
self.rank = 0
def Find(element):
if not element.parent == element:
element.parent = Find(element.parent)
return element.parent
def Union(element_one, element_two):
root_one = Find(element_one)
root_two = Find(element_two)
if root_one == root_two: return
if root_one.rank < root_two.rank:
root_one.parent = root_two
elif root_one.rank > root_two.rank:
root_two.parent = root_one
else:
root_two.parent = root_one
root_one.rank = root_one.rank + 1
union_find_elements = {}
for label in neighbor_label_dicts.keys():
# skip over elements that remain background
if not associated_label_dict[label]: continue
union_find_elements[label] = UnionFindElement(label)
for label in neighbor_label_dicts.keys():
# skip over elements that remain background
if not associated_label_dict[label]: continue
for neighbor_label in neighbor_label_dicts[label]:
# skip over the actual neuron label
if neighbor_label > 0: continue
# merge these two labels together
Union(union_find_elements[label], union_find_elements[neighbor_label])
root_holes_sizes = {}
# go through all labels in the union find data structure and update the hole size for the parent
for label in union_find_elements.keys():
root_label = Find(union_find_elements[label])
# create this hole if it does not already exist
if not root_label.label in root_holes_sizes:
root_holes_sizes[root_label.label] = 0
root_holes_sizes[root_label.label] += hole_sizes[label]
# read in the statistics data to find total volume size
statistics_directory = '{}/statistics'.format(data.TempDirectory())
statistics_filename = '{}/combined-statistics.pickle'.format(statistics_directory)
volume_statistics = ReadPickledData(statistics_filename)
total_volume = volume_statistics['neuronal_volume']
holes = []
small_holes = 0
for root_label in root_holes_sizes.keys():
if root_holes_sizes[root_label] < 5:
small_holes += 1
holes.append(root_holes_sizes[root_label])
# get statistics on the number of holes
nholes = len(holes)
total_hole_volume = sum(holes)
print ('Percent Small: {}'.format(100.0 * small_holes / nholes))
print ('No. Holes: {}'.format(nholes))
print ('Total Volume: {} ({:0.2f}%)'.format(total_hole_volume, 100.0 * total_hole_volume / total_volume))
def EvaluateSkeletons(meta_filename):
data = ReadMetaData(meta_filename)
# make sure a results folder is specified
assert (not data.EvaluationDirectory() == None)
# read in statistics about this data set
statistics_directory = '{}/statistics'.format(data.TempDirectory())
statistics_filename = '{}/combined-statistics.pickle'.format(statistics_directory)
statistics = ReadPickledData(statistics_filename)
label_volumes = statistics['label_volumes']
total_volume = 0
total_thinned_skeleton_length = 0
total_refined_skeleton_length = 0
nlabels = 0
# get the output filename
evaluation_directory = data.EvaluationDirectory()
if not os.path.exists(evaluation_directory):
os.makedirs(evaluation_directory, exist_ok=True)
output_filename = '{}/skeleton-results.txt'.format(evaluation_directory)
fd = open(output_filename, 'w')
for label in sorted(label_volumes):
# read pre-refinement skeleton
thinning_filename = '{}/skeletons/{:016d}.pts'.format(data.TempDirectory(), label)
# skip files that do not exist (no synapses, e.g.)
if not os.path.exists(thinning_filename): continue
thinned_skeletons_global_pts, _ = ReadPtsFile(data, thinning_filename)
thinned_skeletons = thinned_skeletons_global_pts[label]
refined_filename = '{}/skeletons/{:016d}.pts'.format(data.SkeletonOutputDirectory(), label)
refined_skeletons_global_pts, _ = ReadPtsFile(data, refined_filename)
refined_skeletons = refined_skeletons_global_pts[label]
# get the volume and total remaining voxels
volume = label_volumes[label]
thinned_skeleton_length = len(thinned_skeletons)
refined_skeleton_length = len(refined_skeletons)
# update the variables that aggregate all labels
total_volume += volume
total_thinned_skeleton_length += thinned_skeleton_length
total_refined_skeleton_length += refined_skeleton_length
# calculate the percent and reduction of total voxels remaining
thinned_remaining_percent = 100 * thinned_skeleton_length / volume
thinning_reduction_factor = volume / thinned_skeleton_length
refined_remaining_percent = 100 * refined_skeleton_length / thinned_skeleton_length
refinement_reduction_factor = thinned_skeleton_length / refined_skeleton_length
# caluclate the total percent/reduction after all steps
total_skeleton_percent = 100 * refined_skeleton_length / volume
total_skeleton_reduction = volume / refined_skeleton_length
print ('Label: {}'.format(label))
print (' Input Volume: {:10d}'.format(volume))
print (' Topological Thinning: {:10d} ({:05.2f}%) {:10.2f}x'.format(thinned_skeleton_length, thinned_remaining_percent, thinning_reduction_factor))
print (' Skeleton Refinement: {:10d} ({:05.2f}%) {:10.2f}x'.format(refined_skeleton_length, refined_remaining_percent, refinement_reduction_factor))
print (' Total: ({:05.2f}%) {:10.2f}x'.format(total_skeleton_percent, total_skeleton_reduction))
fd.write ('Label: {}\n'.format(label))
fd.write (' Input Volume: {:10d}\n'.format(volume))
fd.write (' Topological Thinning: {:10d} ({:05.2f}%) {:10.2f}x\n'.format(thinned_skeleton_length, thinned_remaining_percent, thinning_reduction_factor))
fd.write (' Skeleton Refinement: {:10d} ({:05.2f}%) {:10.2f}x\n'.format(refined_skeleton_length, refined_remaining_percent, refinement_reduction_factor))
fd.write (' Total: ({:05.2f}%) {:10.2f}x\n'.format(total_skeleton_percent, total_skeleton_reduction))
nlabels += 1
# calculate the percent and reduction of total voxels remaining
thinned_remaining_percent = 100 * total_thinned_skeleton_length / total_volume
thinning_reduction_factor = total_volume / total_thinned_skeleton_length
refined_remaining_percent = 100 * total_refined_skeleton_length / total_thinned_skeleton_length
refinement_reduction_factor = total_thinned_skeleton_length / total_refined_skeleton_length
# caluclate the total percent/reduction after all steps
total_skeleton_percent = 100 * total_refined_skeleton_length / total_volume
total_skeleton_reduction = total_volume / total_refined_skeleton_length
print ('Input Volume: {:10d}'.format(total_volume))
print ('Topological Thinning: {:10d} ({:05.2f}%) {:10.2f}x'.format(total_thinned_skeleton_length, thinned_remaining_percent, thinning_reduction_factor))
print ('Skeleton Refinement: {:10d} ({:05.2f}%) {:10.2f}x'.format(total_refined_skeleton_length, refined_remaining_percent, refinement_reduction_factor))
print ('Total: ({:05.2f}%) {:10.2f}x'.format(total_skeleton_percent, total_skeleton_reduction))
print ('Average Skeleton: {:0.0f}'.format(total_refined_skeleton_length / nlabels))
fd.write ('Input Volume: {:10d}\n'.format(total_volume))
fd.write ('Topological Thinning: {:10d} ({:05.2f}%) {:10.2f}x\n'.format(total_thinned_skeleton_length, thinned_remaining_percent, thinning_reduction_factor))
fd.write ('Skeleton Refinement: {:10d} ({:05.2f}%) {:10.2f}x\n'.format(total_refined_skeleton_length, refined_remaining_percent, refinement_reduction_factor))
fd.write ('Total: ({:05.2f}%) {:10.2f}x\n'.format(total_skeleton_percent, total_skeleton_reduction))
fd.write ('Average Skeleton: {:0.0f}'.format(total_refined_skeleton_length / nlabels))
# close the file
fd.close()