def in_volume_ncoll(points: np.ndarray,
volume: Volume,
n_rays: Optional[int] = 3) -> Sequence[bool]:
"""Use ncollpyde to test if points are within a given volume."""
if isinstance(n_rays, type(None)):
n_rays = 3
if not isinstance(n_rays, (int, np.integer)):
raise TypeError(f'n_rays must be integer, got "{type(n_rays)}"')
if n_rays <= 0:
raise ValueError('n_rays must be > 0')
coll = ncollpyde.Volume(volume.vertices, volume.faces, n_rays=n_rays)
return coll.contains(points)
def _test_single_edge(l1, l2, seg_id, vol):
"""Test single edge.
Parameters
----------
l1, l2 : int | float
Locations of the nodes connected by the edge
seg_id : int
Segment ID of one of the nodes.
vol : cloudvolume.CloudVolume
Returns
-------
bool
"""
# Get the bounding box
bbox = np.array([l1, l2])
bbox = np.array([bbox.min(axis=0), bbox.max(axis=0)]).T
# Get the mesh
mesh = get_mesh(seg_id, bbox=bbox, vol=vol)
# No mesh means that edge is most likely True
if not mesh:
return True
# Prepare raycasting
coll = ncollpyde.Volume(np.array(mesh.vertices, dtype=float, order='C'),
np.array(mesh.faces, dtype=np.int32, order='C'))
# Get intersections
l1 = l1.reshape(1, 3)
l2 = l2.reshape(1, 3)
inter_ix, inter_xyz, is_inside = coll.intersections(l1, l2)
# If not intersections treat this edge as True
if not inter_xyz.any():
return True
return True
def get_radius_ray(swc, mesh, n_rays=20, aggregate='mean', projection='sphere',
fallback='knn'):
"""Extract radii using ray casting.
Parameters
----------
swc : pandas.DataFrame
SWC table
mesh : trimesh.Trimesh
n_rays : int
Number of rays to cast for each node.
aggregate : "mean" | "median" | "max" | "min" | "percentile75"
Function used to aggregate radii for over all intersections
for a given node.
projection : "sphere" | "tangents"
Whether to cast rays in a sphere around each node or in a
circle orthogonally to the node's tangent vector.
fallback : "knn" | None | number
If a point is outside or right on the surface of the mesh
the raycasting will return nonesense results. We can either
ignore those cases (``None``), assign a arbitrary number or
we can fall back to radii from k-nearest-neighbors (``knn``).
Returns
-------
radii : np.ndarray
Corresponds to input coords.
"""
agg_map = {'mean': np.mean, 'max': np.max, 'min': np.min,
'median': np.median, 'percentile75': lambda x: np.percentile(x, 75)}
assert aggregate in agg_map
agg_func = agg_map[aggregate]
assert projection in ['sphere', 'tangents']
assert (fallback == 'knn') or isinstance(fallback, numbers.Number) or isinstance(fallback, type(None))
# Get max dimension of mesh
dim = (swc[['x', 'y', 'z']].max() - swc[['x', 'y', 'z']].min()).values
radius = max(dim)
# Vertices for each point on the circle
points = swc[['x', 'y', 'z']].values
sources = np.repeat(points, n_rays, axis=0)
if projection == 'sphere':
# Repeat points n_rays times
sources = np.repeat(points, n_rays, axis=0)
# Get (random) points on a sphere and scale by radius
targets = fibonacci_sphere(n_rays, randomize=True) * radius
# Reshape to match sources
targets = np.tile(targets,
(points.shape[0], 1))
# Offset onto sources
targets += sources
else:
tangents, normals, binormals = frenet_frames(swc)
v = np.arange(n_rays,
dtype=np.float) / n_rays * 2 * np.pi
all_cx = (radius * -1. * np.tile(np.cos(v), points.shape[0]).reshape((n_rays, points.shape[0]), order='F')).T
cx_norm = (all_cx[:, :, np.newaxis] * normals[:, np.newaxis, :]).reshape(sources.shape)
all_cy = (radius * np.tile(np.sin(v), points.shape[0]).reshape((n_rays, points.shape[0]), order='F')).T
cy_norm = (all_cy[:, :, np.newaxis] * binormals[:, np.newaxis, :]).reshape(sources.shape)
targets = sources + cx_norm + cy_norm
# Initialize ncollpyde Volume
coll = ncollpyde.Volume(mesh.vertices, mesh.faces, validate=False)
# Get intersections: `ix` points to index of line segment; `loc` is the
# x/y/z coordinate of the intersection and `is_backface` is True if
# intersection happened at the inside of a mesh
ix, loc, is_backface = coll.intersections(sources, targets)
# Remove intersections with front faces
# For some reason this reduces the number of intersections to 0 for many
# points
#ix = ix[~is_backface]
#loc = loc[~is_backface]
# Calculate intersection distances
dist = np.sqrt(np.sum((sources[ix] - loc)**2, axis=1))
# Map from `ix` back to index of original point
org_ix = (ix / n_rays).astype(int)
# Split by original index
split_ix = np.where(org_ix[:-1] - org_ix[1:])[0]
split = np.split(dist, split_ix)
# Aggregate over each original ix
final_dist = np.zeros(points.shape[0])
for l, i in zip(split, np.unique(org_ix)):
final_dist[i] = agg_func(l)
if not isinstance(fallback, type(None)):
# See if any needs fixing
inside = coll.contains(points)
is_zero = final_dist == 0
needs_fix = ~inside | is_zero
if any(needs_fix):
if isinstance(fallback, numbers.Number):
final_dist[needs_fix] = fallback
elif fallback == 'knn':
final_dist[needs_fix] = get_radius_kkn(points[needs_fix], mesh, aggregate=aggregate)
return final_dist
def remove_hairs(s, mesh=None, inplace=False):
"""Remove "hairs" that sometimes occurr along the backbone.
Works by finding terminal twigs that consist of only a single node. We will
then remove those that are within line of sight of their parent.
Note that this is currently not used for clean up as it does not work very
well: removes as many correct hairs as genuine small branches.
Parameters
----------
s : skeletor.Skeleton
Skeleton to clean up.
mesh : trimesh.Trimesh, optional
Original mesh (e.g. before contraction). If not provided will
use the mesh associated with ``s``.
max_dist : "auto" | int | float
Maximum Eucledian distance allowed between leaf nodes for them
to be considered for collapsing. If "auto", will use the length
of the longest edge in skeleton as limit.
inplace : bool
If False will make and return a copy of the skeleton. If True,
will modify the `s` inplace.
Returns
-------
SWC : pandas.DataFrame
SWC with line-of-sight twigs removed.
"""
if isinstance(mesh, type(None)):
mesh = s.mesh
# Make a copy of the skeleton
if not inplace:
s = s.copy()
# Find branch points
pcount = s.swc[s.swc.parent_id >= 0].groupby('parent_id').size()
bp = pcount[pcount > 1].index
# Find terminal twigs
twigs = s.swc[~s.swc.node_id.isin(s.swc.parent_id)]
twigs = twigs[twigs.parent_id.isin(bp)]
if twigs.empty:
return s
# Initialize ncollpyde Volume
coll = ncollpyde.Volume(mesh.vertices, mesh.faces, validate=False)
# Remove twigs that aren't inside the volume
twigs = twigs[coll.contains(twigs[['x', 'y', 'z']].values)]
# Generate rays between all pairs and their parents
sources = twigs[['x', 'y', 'z']].values
targets = s.swc.set_index('node_id').loc[twigs.parent_id,
['x', 'y', 'z']].values
# Get intersections: `ix` points to index of line segment; `loc` is the
# x/y/z coordinate of the intersection and `is_backface` is True if
# intersection happened at the inside of a mesh
ix, loc, is_backface = coll.intersections(sources, targets)
# Find pairs of twigs with no intersection - i.e. with line of sight
los = ~np.isin(np.arange(sources.shape[0]), ix)
# To remove: have line of sight
to_remove = twigs[los]
s.swc = s.swc[~s.swc.node_id.isin(to_remove.node_id)].copy()
# Update the mesh map
mesh_map = getattr(s, 'mesh_map', None)
if not isinstance(mesh_map, type(None)):
for t in to_remove.itertuples():
mesh_map[mesh_map == t.node_id] = t.parent_id
# Reindex nodes
s.reindex(inplace=True)
return s
def drop_line_of_sight_twigs(s, mesh=None, max_dist='auto', inplace=False):
"""Collapse twigs that are in line of sight to each other.
Note that this only removes 1 layer of twigs (i.e. only the actual leaf
nodes). Nothing is stopping you from running this function recursively
though.
Also note that this function needs a rework because it does not take
connected components into account and hence collapses things that were
not meant to be connected.
Parameters
----------
s : skeletor.Skeleton
Skeleton to clean up.
mesh : trimesh.Trimesh, optional
Original mesh (e.g. before contraction). If not provided will
use the mesh associated with ``s``.
max_dist : "auto" | int | float
Maximum Eucledian distance allowed between leaf nodes for them
to be considered for collapsing. If "auto", will use the length
of the longest edge in skeleton as limit.
inplace : bool
If False will make and return a copy of the skeleton. If True,
will modify the `s` inplace.
Returns
-------
SWC : pandas.DataFrame
SWC with line-of-sight twigs removed.
"""
# Make a copy of the SWC
if not inplace:
s = s.copy()
# Add distance to parents
s.swc['parent_dist'] = 0
not_root = s.swc.parent_id >= 0
co1 = s.swc.loc[not_root, ['x', 'y', 'z']].values
co2 = s.swc.set_index('node_id').loc[s.swc.loc[not_root, 'parent_id'],
['x', 'y', 'z']].values
s.swc.loc[not_root, 'parent_dist'] = np.sqrt(np.sum((co1 - co2)**2, axis=1))
# If max dist is 'auto', we will use the longest child->parent edge in the
# skeleton as limit
if max_dist == 'auto':
max_dist = s.swc.parent_dist.max()
# Initialize ncollpyde Volume
coll = ncollpyde.Volume(mesh.vertices, mesh.faces, validate=False)
# Find twigs
twigs = s.swc[~s.swc.node_id.isin(s.swc.parent_id)]
# Remove twigs that aren't inside the volume
twigs = twigs[coll.contains(twigs[['x', 'y', 'z']].values)]
# Generate rays between all pairs of twigs
twigs_co = twigs[['x', 'y', 'z']].values
sources = np.repeat(twigs_co, twigs.shape[0], axis=0)
targets = np.tile(twigs_co, (twigs.shape[0], 1))
# Keep track of indices
pairs = np.stack((np.repeat(twigs.node_id, twigs.shape[0]),
np.tile(twigs.node_id, twigs.shape[0]))).T
# If max distance, drop pairs that are too far appart
if max_dist:
d = scipy.spatial.distance.pdist(twigs_co)
d = scipy.spatial.distance.squareform(d)
is_close = d.flatten() <= max_dist
pairs = pairs[is_close]
sources = sources[is_close]
targets = targets[is_close]
# Drop self rays
not_self = pairs[:, 0] != pairs[:, 1]
sources, targets = sources[not_self], targets[not_self]
pairs = pairs[not_self]
# Get intersections: `ix` points to index of line segment; `loc` is the
# x/y/z coordinate of the intersection and `is_backface` is True if
# intersection happened at the inside of a mesh
ix, loc, is_backface = coll.intersections(sources, targets)
# Find pairs of twigs with no intersection - i.e. with line of sight
los = ~np.isin(np.arange(pairs.shape[0]), ix)
# To collapse: have line of sight
to_collapse = pairs[los]
# Group into cluster we need to collapse
G = nx.Graph()
G.add_edges_from(to_collapse)
clusters = nx.connected_components(G)
# When collapsing the clusters, we need to decide which should be the
# winning twig. For this we will use the twig lengths. In theory we ought to
# be more fancy and ask for the distance to the root but that's more
# expensive and it's unclear if it'll work any better.
seg_lengths = twigs.set_index('node_id').parent_dist.to_dict()
to_remove = []
seen = set()
for nodes in clusters:
# We have to be careful not to generate chains here, e.g. A sees B,
# B sees C, C sees D, etc. To prevent this, we will break up these
# clusters into cliques and collapse them by order of size of the
# cliques
for cl in nx.find_cliques(nx.subgraph(G, nodes)):
# Turn into set
cl = set(cl)
# Drop any node that has already been visited
cl = cl - seen
# Skip if less than 2 nodes left in clique
if len(cl) < 2:
continue
# Add nodes to list of visited nodes
seen = seen | cl
# Sort by segment lenghts to find the loosing nodes
loosers = sorted(cl, key=lambda x: seg_lengths[x])[:-1]
to_remove += loosers
# Drop the tips we flagged for removal and the new column we added
s.swc = s.swc[~s.swc.node_id.isin(to_remove)].drop('parent_dist', axis=1)
# Clean up node/vertex order
s.reindex(inplace=True)
return s
def recenter_vertices(s, mesh=None, inplace=False):
"""Move nodes that ended up outside the mesh back inside.
Nodes can end up outside the original mesh e.g. if the mesh contraction
didn't do a good job (most likely because of internal/degenerate faces that
messed up the normals). This function rectifies this by snapping those nodes
nodes back to the closest vertex and then tries to move them into the
mesh's center. That second step is not guaranteed to work but at least you
won't have any more nodes outside the mesh.
Please note that if connected (!) nodes end up on the same position (i.e
because they snapped to the same vertex), we will collapse them.
Parameters
----------
s : skeletor.Skeleton
mesh : trimesh.Trimesh
Original mesh.
inplace : bool
If False will make and return a copy of the skeleton. If True,
will modify the `s` inplace.
Returns
-------
SWC : pandas.DataFrame
SWC with line-of-sight twigs removed.
"""
if isinstance(mesh, type(None)):
mesh = s.mesh
# Copy skeleton
if not inplace:
s = s.copy()
# Find nodes that are outside the mesh
coll = ncollpyde.Volume(mesh.vertices, mesh.faces, validate=False)
outside = ~coll.contains(s.vertices)
# Skip if all inside
if not any(outside):
return s
# For each outside find the closest vertex
tree = scipy.spatial.cKDTree(mesh.vertices)
# Find nodes that are right on top of original vertices
dist, ix = tree.query(s.vertices[outside])
# We don't want to just snap them back to the closest vertex but try to find
# the center. For this we will:
# 1. Move each vertex inside the mesh by just a bit
# 2. Cast a ray along the vertices' normals and find the opposite sides of the mesh
# 3. Calculate the distance
# Get the closest vertex...
closest_vertex = mesh.vertices[ix]
# .. and offset the vertex positions by just a bit so they "should" be
# inside the mesh. In reality that doesn't always happen if the mesh is not
# watertight
vnormals = mesh.vertex_normals[ix]
sources = closest_vertex - vnormals
# Prepare rays to cast
targets = sources - vnormals * 1e4
# Cast rays
ix, loc, is_backface = coll.intersections(sources, targets)
# If no collisions
if not any(loc):
return s
# Get half-vector
halfvec = np.zeros(sources.shape)
halfvec[ix] = (loc - closest_vertex[ix]) / 2
# Offset vertices
final_pos = closest_vertex + halfvec
# Keep only those that are properly inside the mesh and fall back to the
# closest vertex if that's not the case
now_inside = coll.contains(final_pos)
final_pos[~now_inside] = closest_vertex[~now_inside]
# Replace coordinates
s.swc.loc[outside, 'x'] = final_pos[:, 0]
s.swc.loc[outside, 'y'] = final_pos[:, 1]
s.swc.loc[outside, 'z'] = final_pos[:, 2]
# At this point we may have nodes that snapped to the same vertex and
# therefore end up at the same position. We will collapse those nodes
# - but only if they are actually connected!
# First find duplicate locations
u, i, c = np.unique(s.vertices, return_counts=True, return_inverse=True, axis=0)
# If any coordinates have counter higher 1
if c.max() > 1:
rewire = {}
# Find out which unique coordinates are duplicated
dupl = np.where(c > 1)[0]
# Go over each duplicated coordinate
for ix in dupl:
# Find the nodes on this duplicate coordinate
node_ix = np.where(i == ix)[0]
# Get their edges
edges = s.edges[np.all(np.isin(s.edges, node_ix), axis=1)]
# We will work on the graph to collapse nodes sequentially A->B->C
G = nx.DiGraph()
G.add_edges_from(edges)
for cc in nx.connected_components(G.to_undirected()):
# Root is the node without any outdegree in this subgraph
root = [n for n in cc if G.out_degree[n] == 0][0]
# We don't want to collapse into `root` because it's not actually
# among the nodes with the same coordinates but rather the "last"
# nodes parent
clps_into = next(G.predecessors(root))
# Keep track of how we need to rewire
rewire.update({c: clps_into for c in cc if c not in {root, clps_into}})
# Only mess with the skeleton if there were nodes to be merged
if rewire:
# Rewire
s.swc['parent_id'] = s.swc.parent_id.map(lambda x: rewire.get(x, x))
# Drop nodes that were collapsed
s.swc = s.swc.loc[~s.swc.node_id.isin(rewire)]
# Update mesh map
if not isinstance(s.mesh_map, type(None)):
s.mesh_map = [rewire.get(x, x) for x in s.mesh_map]
# Reindex to make vertex IDs continous again
s.reindex(inplace=True)
# This prevents future SettingsWithCopy Warnings:
if not inplace:
s.swc = s.swc.copy()
return s
def by_tangent_ball(mesh):
"""Skeletonize a mesh by finding the maximal tangent ball.
This algorithm casts a ray from every mesh vertex along its inverse normals
(requires `ncollpyde`). It then creates a sphere that is tangent to the
vertex and to where the ray hit the inside of a face on the opposite side.
Next it drops spheres that overlap with another, larger sphere. Modified
from [1].
The method works best on smooth meshes and is rather sensitive to errors in
the mesh such as incorrect normals (see `skeletor.pre.fix_mesh`), internal
faces, noisy surface (try smoothing or downsampling) or holes in the mesh.
Parameters
----------
mesh : mesh obj
The mesh to be skeletonize. Can an object that has
``.vertices`` and ``.faces`` properties (e.g. a
trimesh.Trimesh) or a tuple ``(vertices, faces)`` or a
dictionary ``{'vertices': vertices, 'faces': faces}``.
Returns
-------
skeletor.Skeleton
Holds results of the skeletonization and enables quick
visualization.
Examples
--------
>>> import skeletor as sk
>>> mesh = sk.example_mesh()
>>> fixed = sk.pre.fix_mesh(mesh, fix_normals=True, remove_disconnected=10)
>>> skel = sk.skeletonize.by_tangent_ball(fixed)
References
----------
[1] Ma, J., Bae, S.W. & Choi, S. 3D medial axis point approximation using
nearest neighbors and the normal field. Vis Comput 28, 7–19 (2012).
https://doi.org/10.1007/s00371-011-0594-7
"""
mesh = make_trimesh(mesh, validate=False)
# Generate the KD tree
tree = scipy.spatial.cKDTree(mesh.vertices)
dist = tree.query(mesh.vertices, k=2)[0][:, 1]
centers = np.zeros(mesh.vertices.shape)
radii = np.zeros(mesh.vertices.shape[0])
coll = ncollpyde.Volume(mesh.vertices, mesh.faces, validate=False)
sources = mesh.vertices - mesh.vertex_normals * 0.01
targets = mesh.vertices - mesh.vertex_normals * (dist.max() * 10)
ix, loc, is_backface = coll.intersections(sources, targets)
# Now we need to invalidate centers
intersects = np.zeros(mesh.vertices.shape[0]).astype(bool)
intersects[ix[is_backface]] = True
centers[ix] = mesh.vertices[ix] + (loc - mesh.vertices[ix]) / 2
radii[ix] = np.linalg.norm(loc - mesh.vertices[ix], axis=1) / 2
# Now we need to post processing
inv = np.zeros(mesh.vertices.shape[0]).astype(bool)
# Invalidate vertices that didn't intersect
inv[~intersects] = True
# Now invalidate any ball that is outside the mesh
inv[~coll.contains(centers)] = True
# Find tangent balls that are fully contained in another tangent ball
# (those are not maximal inscribed)
original_ind = np.arange(mesh.vertices.shape[0])
while True:
tree2 = scipy.spatial.cKDTree(centers[~inv])
# For any not-yet-invalidated center find the closest other center
dist, ix = tree2.query(centers[~inv], k=2)
# Drop self-hits
ix, dist = ix[:, 1], dist[:, 1]
# In radius
in_radius = dist < radii[~inv]
# Stop if no more overlapping pairs
if not in_radius.any():
break
# Collect radii to determine which of the overlapping ball survives
pair_rad = np.vstack((radii[~inv][in_radius],
radii[~inv][ix[in_radius]])).T
pair_ix = np.vstack((original_ind[~inv][in_radius],
original_ind[~inv][ix[in_radius]])).T
# Invalidate the loosers
looses = np.argmax(pair_rad, axis=1)
looser_ix = np.unique(pair_ix[np.arange(pair_ix.shape[0]), looses])
inv[looser_ix] = True
# Now we need to collapse nodes into the remaining centers
G = ig.Graph(n=mesh.vertices.shape[0],
edges=mesh.edges_unique,
directed=False)
# Make sure that every connected component has at least one valid target
for cc in G.clusters():
if not np.isin(cc, original_ind[~inv]).any():
inv[cc[0]] = False
centers[cc[0]] = mesh.vertices[cc[0]]
# For each invalidated vertex, find the closest vertex that is still valid
# This works on unweighted edges but should be good enough - way faster
# than a proper path search for sure
pairs = find_closest(G, sources=original_ind[inv],
targets=original_ind[~inv])
# Generate a mesh vertex to skeleton node map
mesh_map = original_ind.copy()
mesh_map[pairs[:, 0]] = pairs[:, 1]
# Renumber the vertices from 0 -> N_vertices
uni, ind, mesh_map = np.unique(mesh_map, return_inverse=True, return_index=True)
# Make sure centers and radii match the new order
centers = centers[uni]
radii = radii[uni]
# Contract vertices to nodes according to the mesh
G.contract_vertices(mesh_map, combine_attrs=None)
# This only drops duplicate and self-loop edges
G = G.simplify()
# Generate weights between remaining centers
el = np.array(G.get_edgelist())
weights = np.linalg.norm(centers[el[:, 0]] - centers[el[:, 1]], axis=1)
# Generate hierarchical tree
tree = G.spanning_tree(weights=weights)
# Create a directed acyclic and hierarchical graph
G_nx = edges_to_graph(edges=np.array(tree.get_edgelist()),
nodes=np.arange(0, len(G.vs)),
fix_tree=True,
drop_disconnected=False)
# Generate the SWC table
swc = make_swc(G_nx, coords=centers, reindex=False)
swc['radius'] = radii[swc.node_id.values]
_, new_ids = reindex_swc(swc, inplace=True)
# Update vertex to node ID map
mesh_map = np.array([new_ids[n] for n in mesh_map])
return Skeleton(swc=swc, mesh=mesh, mesh_map=mesh_map,
method='tangent_ball')
def collapse_nodes(A, B, limit=1, base_neuron=None, mesh=None):
"""Merge neuron A into neuron(s) B creating a union of both.
This implementation uses edge contraction on the neurons' graph to ensure
maximum connectivity. Only works if the fragments collectively form a
continuous tree (i.e. you must be certain that they partially overlap).
Parameters
----------
A : CatmaidNeuron
Neuron to be collapsed into neurons B.
B : CatmaidNeuronList
Neurons to collapse neuron A into.
limit : int, optional
Max distance [microns] for nearest neighbour search.
base_neuron : skeleton ID | CatmaidNeuron, optional
Neuron from B to use as template for union. If not
provided, the first neuron in the list is used as
template!
mesh : navis.Volume | navis.MeshNeuron | mesh-like object
If provided, will use the mesh to check if nodes are
in line of sight to each other before collapsing them.
Returns
-------
core.CatmaidNeuron
Union of all input neurons.
new_edges : pandas.DataFrame
Subset of the ``.nodes`` table that represent newly
added edges.
collapsed_nodes : dict
Map of collapsed nodes::
NodeA -collapsed-into-> NodeB
"""
if isinstance(A, navis.NeuronList):
if len(A) == 1:
A = A[0]
else:
A = navis.stitch_neurons(A, method="NONE")
if not isinstance(A, navis.TreeNeuron):
raise TypeError('`A` must be a TreeNeuron, got "{}"'.format(type(A)))
if isinstance(B, navis.TreeNeuron):
B = navis.NeuronList(B)
if not isinstance(B, navis.NeuronList):
raise TypeError('`B` must be a NeuronList, got "{}"'.format(type(B)))
if isinstance(base_neuron, type(None)):
base_neuron = B[0]
# This is just check on the off-chance that skeleton IDs are not unique
# (e.g. if neurons come from different projects) -> this is relevant because
# we identify the master ("base_neuron") via it's skeleton ID
skids = [n.id for n in B + A]
if len(skids) > len(set(skids)):
raise ValueError(
'Duplicate skeleton IDs found. Try manually assigning '
'unique skeleton IDs.')
# Convert distance threshold from microns to nanometres
limit *= 1000
# Before we start messing around, let's make sure we can keep track of
# the origin of each node
for n in B + A:
n.nodes['origin_skeletons'] = n.id
# First make a weak union by simply combining the node tables
B.neurons = sorted(B.neurons, key=lambda x: {base_neuron: 0}.get(x, 2))
union_simple = navis.stitch_neurons(B + A, method='NONE', master='FIRST')
# Check for duplicate node IDs
if any(union_simple.nodes.node_id.duplicated()):
raise ValueError('Duplicate node IDs found.')
# Find nodes in A to be merged into B
tree = scipy.spatial.cKDTree(data=B.nodes[['x', 'y', 'z']].values)
# For each node in A get the nearest neighbor in B
coords = A.nodes[['x', 'y', 'z']].values
nn_dist, nn_ix = tree.query(coords, k=1, distance_upper_bound=limit)
# Find nodes that are close enough to collapse
collapsed = A.nodes.loc[nn_dist <= limit].node_id.values
clps_into = B.nodes.iloc[nn_ix[nn_dist <= limit]].node_id.values
# If we have a mesh, check if those collapsed nodes are in sight of each
# other
if mesh:
import ncollpyde
coll = ncollpyde.Volume(mesh.vertices, mesh.faces)
# Produce start and end coordinates for the to collapse nodes
starts = A.nodes.set_index('node_id').loc[collapsed,
['x', 'y', 'z']].values
ends = B.nodes.set_index('node_id').loc[clps_into,
['x', 'y', 'z']].values
# Check if the line between start and end intersects the mesh
intersects, _, _ = coll.intersections(starts, ends)
# Indices that show up in `intersects` cross a membrane
# -> we need to invert this to get those nodes that don't
not_intersects = ~np.isin(np.arange(starts.shape[0]), intersects)
# Keep only collapses that don't intersect
collapsed = collapsed[not_intersects]
clps_into = clps_into[not_intersects]
# Generate a map of which node in A is to be collapsed into which node in B
clps_map = {n1: n2 for n1, n2 in zip(collapsed, clps_into)}
# The fastest way to collapse is to work on the edge list
E = nx.to_pandas_edgelist(union_simple.graph)
# Keep track of which edges were collapsed -> we will use this as weight
# later on to prioritize existing edges over newly generated ones
E['is_new'] = 1
source_in_B = E.source.isin(B.nodes.node_id.values)
target_in_B = E.target.isin(B.nodes.node_id.values)
E.loc[source_in_B | target_in_B, 'is_new'] = 0
# Now map collapsed nodes onto the nodes they collapsed into
E['target'] = E.target.map(lambda x: clps_map.get(x, x))
E['source'] = E.source.map(lambda x: clps_map.get(x, x))
# Make sure no self loops after collapsing. This happens if two adjacent
# nodes collapse onto the same target node
E = E[E.source != E.target]
# Remove duplicates. This happens e.g. when two adjaceny nodes merge into
# two other adjaceny nodes: A->B C->D ----> A/B->C/D
# By sorting first, we make sure original edges are kept first
E.sort_values('is_new', ascending=True, inplace=True)
# Because edges may exist in both directions (A->B and A<-B) we have to
# generate a column that's agnostic to directionality using frozensets
E['frozen_edge'] = E[['source', 'target']].apply(frozenset, axis=1)
E.drop_duplicates(['frozen_edge'], keep='first', inplace=True)
# Regenerate graph from these new edges
G = nx.Graph()
G.add_weighted_edges_from(E[['source', 'target',
'is_new']].values.astype(int))
# At this point there might still be disconnected pieces -> we will create
# separate neurons for each tree
props = union_simple.nodes.loc[union_simple.nodes.node_id.isin(
G.nodes)].set_index('node_id')
nx.set_node_attributes(G, props.to_dict(orient='index'))
fragments = []
for n in nx.connected_components(G):
c = G.subgraph(n)
tree = nx.minimum_spanning_tree(c)
fragments.append(
navis.graph.nx2neuron(tree,
name=base_neuron.name,
id=base_neuron.id))
fragments = navis.NeuronList(fragments)
if len(fragments) > 1:
print('Union incomplete - watch out for disconnected fragments!')
# Now heal those fragments using a minimum spanning tree
union = navis.stitch_neurons(*fragments, method='ALL')
else:
union = fragments[0]
# Reroot to base neuron's root
union.reroot(base_neuron.root[0], inplace=True)
# Add tags back on
if union_simple.has_tags:
if not union.has_tags:
union.tags = {}
union.tags.update(union_simple.tags)
# Add connectors back on
union.connectors = union_simple.connectors.drop_duplicates(
subset='connector_id').copy()
union.connectors.loc[:, 'node_id'] = union.connectors.node_id.map(
lambda x: clps_map.get(x, x))
# Find the newly added edges (existing edges should not have been modified
# - except for changing direction due to reroot)
# The basic logic here is that new edges were only added between two
# previously separate skeletons, i.e. where the skeleton ID changes between
# parent and child node
node2skid = union_simple.nodes.set_index(
'node_id').origin_skeletons.to_dict()
union.nodes['parent_skeleton'] = union.nodes.parent_id.map(node2skid)
new_edges = union.nodes[
union.nodes.origin_skeletons != union.nodes.parent_skeleton]
# Remove root edges
new_edges = new_edges[~new_edges.parent_id.isnull()]
return union, new_edges, clps_map