from __future__ import annotations
from typing import ClassVar, TYPE_CHECKING
if TYPE_CHECKING:
from phenotypic._core._image import Image
import numpy as np
import pandas as pd
import scipy.ndimage as ndi
from skimage.measure import regionprops_table
from phenotypic.abc_ import MeasureFeatures
from phenotypic.schema import OBJECT
from phenotypic.schema import BBOX
[docs]
class MeasureBounds(MeasureFeatures):
"""Extract bounding box coordinates and centroids of detected colonies.
Compute the axis-aligned bounding box and centroid (geometric and
intensity-weighted) for each detected colony. These spatial
measurements form the foundation for region-of-interest extraction,
grid alignment assessment, and neighbor-distance calculations.
Returns:
pd.DataFrame: Object-level spatial data with columns:
- Label: unique object identifier.
- CenterRR, CenterCC: geometric centroid coordinates.
- IntensityWeightedCenterRR, IntensityWeightedCenterCC:
intensity-weighted centroid coordinates (skimage
``centroid_weighted``).
- DistWeightedCenterRR, DistWeightedCenterCC: per-object
DT-weighted centroid (Sum(dt*pos)/Sum(dt)) — robust to thin
filamentous extensions (low DT weight) and to budding doublets
(the two lobes balance to the neck).
- MinRR, MinCC: top-left corner of bounding box.
- MaxRR, MaxCC: bottom-right corner of bounding box.
Best For:
- Computing centroids for aligning colonies to expected grid
positions in arrayed assays.
- Extracting region-of-interest crops for downstream intensity,
color, or texture analysis.
- Assessing colony positioning relative to plate edges or well
boundaries.
Consider Also:
- :class:`MeasureShape` for full morphological metrics built on
top of bounding box data.
- :class:`MeasureGridLinRegStats` for regression-based grid
alignment quality using centroid positions.
- :class:`MeasureGridSpatial` for neighbor distance calculations
using bounding boxes.
See Also:
:doc:`/tutorials/notebooks/07_measuring_and_exporting` for a
walkthrough of measuring and exporting colony data.
"""
_measurement_infoclass: ClassVar[type] = BBOX
def _operate(self, image: Image) -> pd.DataFrame:
objmap = image.objmap[:]
# gray is stored float32; upcast for the intensity-weighted centroid
# accumulation so values match the historical float64 layer.
gray = image.gray[:].astype(np.float64, copy=False)
results = pd.DataFrame(
data=regionprops_table(
label_image=objmap,
intensity_image=gray,
properties=["label", "centroid", "bbox", "centroid_weighted"]
)
).rename(
columns={
"label" : OBJECT.LABEL,
"centroid-0" : str(BBOX.CENTER_RR),
"centroid-1" : str(BBOX.CENTER_CC),
"centroid_weighted-0": str(BBOX.INTENSITY_WEIGHTED_CENTER_RR),
"centroid_weighted-1": str(BBOX.INTENSITY_WEIGHTED_CENTER_CC),
"bbox-0" : str(BBOX.MIN_RR),
"bbox-1" : str(BBOX.MIN_CC),
"bbox-2" : str(BBOX.MAX_RR),
"bbox-3" : str(BBOX.MAX_CC),
}
)
# Per-object DT-weighted centroid (Sum(dt*pos)/Sum(dt)). Robust to thin
# filamentous extensions (low DT weight keeps the center on-body) and to
# budding doublets (the two lobes' DT mass balances to the neck), unlike a
# DT-argmax which would snap onto a single lobe.
#
# Vectorized via global EDT + ndi.center_of_mass. Inter-object
# boundaries are zeroed before the EDT so touching colonies don't
# bleed into each other.
#
# Memory note: the global EDT allocates a full-image float64 array
# (~128 MB at 4000x4000, ~288 MB at 6000x6000). If this becomes a
# bottleneck on large plates, switch to a per-object loop using
# scipy.ndimage.find_objects + per-bbox distance_transform_edt; the
# pattern at _measure_symmetric_zones.py:237-239 is the precedent.
nonzero = objmap > 0
if nonzero.any():
max_label_value = np.iinfo(objmap.dtype).max
masked_for_min = np.where(nonzero, objmap, max_label_value)
nbr_max = ndi.maximum_filter(objmap, size=3)
nbr_min_nz = ndi.minimum_filter(masked_for_min, size=3)
inter_boundary = nonzero & (
(nbr_max > objmap) | (nbr_min_nz < objmap)
)
binary = nonzero & ~inter_boundary
dt = ndi.distance_transform_edt(binary)
labels = np.unique(objmap[nonzero])
# DT-weighted centroid (Sum(dt*pos)/Sum(dt)) — robust to filament hyphae
# (low DT weight) AND budding doublets (two lobes balance to the neck),
# replacing the former DT-argmax which snapped onto a single lobe.
positions = np.asarray(
ndi.center_of_mass(dt, labels=objmap, index=labels), dtype=float
)
# Degenerate guard: objects whose pixels were all zeroed as inter-object
# boundary have zero DT mass -> center_of_mass returns NaN. Fall back to
# the unweighted (geometric) mask centroid for those labels.
nan_rows = np.isnan(positions).any(axis=1)
if nan_rows.any():
geom = np.asarray(
ndi.center_of_mass(nonzero.astype(float), labels=objmap, index=labels),
dtype=float,
)
positions[nan_rows] = geom[nan_rows]
dist_df = pd.DataFrame({
OBJECT.LABEL : labels,
str(BBOX.DIST_WEIGHTED_CENTER_RR): positions[:, 0],
str(BBOX.DIST_WEIGHTED_CENTER_CC): positions[:, 1],
})
results = results.merge(dist_df, on=OBJECT.LABEL, how="left")
else:
results[str(BBOX.DIST_WEIGHTED_CENTER_RR)] = np.array([], dtype=float)
results[str(BBOX.DIST_WEIGHTED_CENTER_CC)] = np.array([], dtype=float)
return results
MeasureBounds.__doc__ = BBOX.append_rst_to_doc(MeasureBounds)
