File size: 6,882 Bytes
08764e9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 | """Root-canal label cleaning (Stage 0).
Implements the three-step cleaning from the report:
1. connected components (26-connectivity)
2. remove specks smaller than a voxel threshold (NOT keep-largest, so real
multi-canals in molars survive)
3. morphological closing to bridge artifact-induced gaps (+ optional hole fill)
Also returns before/after statistics for the QC table.
"""
import numpy as np
from scipy import ndimage as ndi
from skimage import morphology
_FULL26 = np.ones((3, 3, 3), dtype=int)
def _ball(radius):
return morphology.ball(int(radius)) if radius and radius > 0 else None
def clean_binary(mask, speck_min_voxels=30, closing_radius=1, fill_holes=True):
"""Clean a single binary structure. Returns (cleaned_mask, stats)."""
mask = mask.astype(bool)
lab, n = ndi.label(mask, structure=_FULL26)
stats = dict(n_components_before=int(n),
voxels_before=int(mask.sum()))
if n == 0:
stats.update(n_components_after=0, voxels_after=0, removed_specks=0,
largest_ratio=0.0)
return mask, stats
sizes = ndi.sum(np.ones_like(lab), lab, index=np.arange(1, n + 1))
largest = float(sizes.max())
stats["largest_ratio"] = float(largest / max(mask.sum(), 1))
keep = np.zeros_like(mask)
removed = 0
for i, s in enumerate(sizes, start=1):
if s >= speck_min_voxels:
keep |= (lab == i)
else:
removed += 1
# safety: if threshold removed everything, fall back to largest component
if not keep.any():
keep = (lab == (1 + int(np.argmax(sizes))))
if closing_radius and closing_radius > 0:
keep = ndi.binary_closing(keep, structure=_ball(closing_radius))
if fill_holes:
keep = ndi.binary_fill_holes(keep)
lab2, n2 = ndi.label(keep, structure=_FULL26)
stats.update(n_components_after=int(n2),
voxels_after=int(keep.sum()),
removed_specks=int(removed))
return keep.astype(bool), stats
def _nearest_body_field(body_map):
"""For every voxel, the id of the spatially nearest tooth body.
Propagates body ids into the pulp space / background via the EDT, so a canal
voxel sitting inside a tooth resolves to that tooth's id regardless of how the
raw canal label was numbered."""
inds = ndi.distance_transform_edt(body_map == 0, return_indices=True)[1]
return body_map[tuple(inds)]
def build_instances(label_xyz, cfg):
"""From the raw multi-label volume build per-tooth (body, canal) instances.
Returns dict instance_id(1..28) -> {body: bool[x,y,z], canal: bool[x,y,z],
tooth_solid: bool[x,y,z]} plus an aggregate cleaning report list.
pair_mode (label_scheme.pair_mode):
'geometric' (default) -- assign each canal component to the tooth BODY it
physically sits inside (nearest-body), ignoring the raw canal id.
Robust to mis-numbered / offset-shifted annotations (the cause of the
canal<->tooth ID mismatches found in QC, e.g. cases 032/035/013/014).
'offset' -- legacy: pair canal label i with body label i+pair_offset.
"""
ls = cfg["label_scheme"]
pp = cfg["preprocess"]
off = ls["pair_offset"]
mode = ls.get("pair_mode", "geometric")
report = []
instances = {}
if mode == "offset":
for i in range(ls["canal_lo"], ls["canal_hi"] + 1):
canal_raw = (label_xyz == i)
body_raw = (label_xyz == (i + off))
if not canal_raw.any() and not body_raw.any():
continue
canal, st = clean_binary(canal_raw,
speck_min_voxels=pp["speck_min_voxels"],
closing_radius=pp["closing_radius"],
fill_holes=pp["fill_holes"])
st["instance"] = i
report.append(st)
tooth_solid = ndi.binary_fill_holes((body_raw | canal).astype(bool))
canal = canal & tooth_solid
instances[i] = dict(body=body_raw, canal=canal, tooth_solid=tooth_solid)
return instances, report
# ---------- geometric re-pairing (default) ----------
# 1) body label map (id = canal-equivalent index i; body lives at label i+off)
body_map = np.zeros(label_xyz.shape, dtype=np.int16)
body_ids = []
for i in range(ls["canal_lo"], ls["canal_hi"] + 1):
bmask = (label_xyz == (i + off))
if bmask.any():
body_map[bmask] = i
body_ids.append(i)
if not body_ids:
return instances, report
# 2) clean the WHOLE canal volume once, then split into components
canal_all = np.zeros(label_xyz.shape, dtype=bool)
for i in range(ls["canal_lo"], ls["canal_hi"] + 1):
canal_all |= (label_xyz == i)
canal_all, _ = clean_binary(canal_all,
speck_min_voxels=pp["speck_min_voxels"],
closing_radius=pp["closing_radius"],
fill_holes=pp["fill_holes"])
# 3) nearest-body id for every voxel
nearest = _nearest_body_field(body_map)
# 4) assign each canal component to the tooth it sits inside (majority vote)
cc, ncc = ndi.label(canal_all, structure=_FULL26)
assigned = {i: np.zeros(label_xyz.shape, dtype=bool) for i in body_ids}
n_reassigned = 0
n_orphan = 0
for k in range(1, ncc + 1):
comp = (cc == k)
votes = nearest[comp]
votes = votes[votes > 0]
if votes.size == 0:
n_orphan += int(comp.sum())
continue
vals, counts = np.unique(votes, return_counts=True)
tgt = int(vals[counts.argmax()])
conf = float(counts.max() / votes.size)
if conf < 0.5 or tgt not in assigned: # ambiguous -> drop as orphan
n_orphan += int(comp.sum())
continue
assigned[tgt] |= comp
# would the legacy offset rule have put this elsewhere? (diagnostic only)
raw_here = label_xyz[comp]
raw_here = raw_here[(raw_here >= ls["canal_lo"]) & (raw_here <= ls["canal_hi"])]
if raw_here.size:
raw_id = int(np.bincount(raw_here).argmax())
if raw_id != tgt:
n_reassigned += int(comp.sum())
# 5) materialize instances
for i in body_ids:
body_raw = (label_xyz == (i + off))
canal = assigned[i]
tooth_solid = ndi.binary_fill_holes((body_raw | canal).astype(bool))
canal = canal & tooth_solid
instances[i] = dict(body=body_raw, canal=canal, tooth_solid=tooth_solid)
report.append(dict(pairing="geometric", n_components=int(ncc),
voxels_reassigned=int(n_reassigned),
voxels_orphaned=int(n_orphan),
n_instances=len(body_ids)))
return instances, report
|