from typing import Literal
import numpy as np
from phenotypic.abc_ import PrefabPipeline
from phenotypic.enhance import EnhanceLocalContrast, MedianFilter, EnhanceBlockMatch
from phenotypic.detect import RoundPeaksDetector
from phenotypic.correction import GridAligner
from phenotypic.refine import ReduceSectionsByLine, GridOversizedObjectRemover
from phenotypic.refine import RemoveBorderObjects, SmallObjectRemover
from phenotypic.refine import MaskFill, MaskOpening
from phenotypic.measure import (
MeasureIntensity,
MeasureShape,
MeasureTexture,
MeasureColor,
)
[docs]
class HeavyRoundPeaksPipeline(PrefabPipeline):
"""Detect and measure round colonies using peak detection with full refinement.
An extended version of :class:`RoundPeaksPipeline` that adds BM3D denoising,
EnhanceLocalContrast contrast enhancement, morphological refinement, grid alignment, and a
second detection pass for improved accuracy on challenging plates.
Steps:
1. EnhanceBlockMatch — high-quality denoising
2. EnhanceLocalContrast — boost local contrast
3. MedianFilter — remove residual speckle
4. RoundPeaksDetector — first detection pass
5. MaskOpening, RemoveBorderObjects, SmallObjectRemover, MaskFill — cleanup
6. GridOversizedObjectRemover, ReduceSectionsByLine — grid refinement
7. GridAligner — straighten the grid
8. RoundPeaksDetector — second detection pass after alignment
9. RemoveBorderObjects, SmallObjectRemover, MaskFill — final cleanup
10. ReduceSectionsByLine — final grid refinement
Measurements: MeasureShape, MeasureColor, MeasureIntensity, MeasureTexture.
Best For:
- Round colonies on grid plates that need thorough refinement.
- Noisy or low-contrast plates where :class:`RoundPeaksPipeline`
produces too many artifacts.
- Plates with slight rotation that benefits from grid alignment
between detection passes.
Consider Also:
- :class:`RoundPeaksPipeline` for a faster, lighter version when
plates are clean and well-lit.
- :class:`HeavyOtsuPipeline` for general-purpose Otsu-based detection.
See Also:
:doc:`/tutorials/notebooks/08_using_prefab_pipelines` for a visual
comparison of prefab pipelines.
"""
[docs]
def __init__(
self,
# Preprocessing / enhancement
bm3d_sigma: float = 0.02,
bm3d_block_size: int = 8,
bm3d_stage_arg: Literal["all_stages", "hard_thresholding"] = "all_stages",
bm3d_clip: bool = True,
clahe_kernel_size: int | None = None,
clahe_clip_limit: float = 0.01,
median_mode: Literal[
"nearest", "reflect", "constant", "mirror", "wrap"] = "nearest",
median_shape: Literal["disk", "square", "diamond"] = "diamond",
median_radius: int = 5,
median_cval: float = 0.0,
# Detection settings
detector_thresh_method: Literal[
"otsu", "mean", "local", "triangle", "minimum", "isodata", "li"
] = "otsu",
detector_subtract_background: bool = True,
detector_remove_noise: bool = True,
detector_footprint_width: int = 6,
detector_noise_radius: int = 1,
detector_smoothing_sigma: float = 2.0,
detector_min_peak_distance: int | None = None,
detector_peak_prominence: float | None = None,
detector_edge_refinement: bool = True,
detector_selection_mode: Literal[
"dominant", "centered", "regularized"] = "dominant",
detector_split_merged: bool = True,
# Grid alignment
grid_aligner_axis: int = 0,
grid_aligner_mode: str = "edge",
# Morphology / refinement
mask_opener_footprint: Literal[
"auto", "disk", "square", "diamond"] | np.ndarray | None = "auto",
mask_opener_width: int = 5,
mask_opener_n_iter: int = 1,
border_remover_size: int | float = 1,
small_object_min_size: int = 50,
mask_fill_structure: np.ndarray | None = None,
mask_fill_origin: int = 0,
# Measurements
texture_scale: int | list[int] = 5,
texture_quant_lvl: Literal[8, 16, 32, 64] = 32,
texture_enhance: bool = False,
texture_warn: bool = False,
color_white_chroma_max: float = 4.0,
color_chroma_min: float = 8.0,
color_include_XYZ: bool = False,
# Pipeline bookkeeping
benchmark: bool = False,
verbose: bool = False,
) -> None:
"""
Represents an image processing pipeline for analyzing microbe colonies on solid media agar.
The pipeline includes preprocessing, detection, morphological refinement, and measurement
steps.
Attributes:
bm3d_sigma: Controls the degree of noise reduction during BM3D denoising. Lower values
retain more fine details, which might preserve subtle colony textures. Higher values
remove more noise but may blur colony edges, affecting detection accuracy.
bm3d_block_size: Block size for BM3D denoising. Larger blocks capture more spatial
context for noise estimation but increase computation time.
bm3d_stage_arg: Specifies the stage of BM3D denoising. "all_stages" applies more
comprehensive denoising, potentially enhancing signal uniformity but may result
in detail loss. "hard_thresholding" retains more high-frequency details but may
leave more background noise intact.
bm3d_clip: Whether to clip denoised values to the [0, 1] range. Disabling may
preserve subtle intensity variations but can produce out-of-range values.
clahe_kernel_size: Determines the size of the kernel used for local contrast enhancement
via EnhanceLocalContrast. Larger sizes improve contrast over broader areas, but may over-amplify
large background variations. Smaller sizes enhance localized details but may
introduce noise.
clahe_clip_limit: Contrast clipping limit for EnhanceLocalContrast. Lower values produce more subtle
enhancement; higher values amplify local contrast more aggressively, which can
over-enhance noise.
median_mode: Boundary handling mode for median filtering. Controls how pixel values
are extrapolated at image edges.
median_shape: Defines the morphological shape ("disk", "square", "diamond") used for
median filtering. The choice impacts how texture and artifacts are smoothed.
For instance, "disk" may preserve radial features, whereas "square" provides
edge-focused filtering.
median_radius: Dictates the width for median filtering. Smaller values enhance fine
textural differences, whereas larger radii smooth broader regions, potentially
affecting the precise detection of small colonies.
median_cval: Constant value used to pad image borders when median_mode is "constant".
detector_thresh_method: Specifies the thresholding method for binary segmentation.
"otsu" (default) applies global thresholding, "mean" uses mean-based threshold,
"local" adapts to background variations, "li" uses Li's iterative minimum
cross-entropy method, and "triangle", "minimum", "isodata" offer alternative
thresholding strategies.
detector_subtract_background: Toggles white tophat background subtraction before
thresholding. Enabling this helps standardize varying lighting or agar density but
may also obscure genuine gradients or subtle ring colonies.
detector_remove_noise: Sets whether morphological opening is applied to remove small
noise artifacts. True ensures a cleaner output but may falsely discard tiny colonies.
False retains all details, which can increase false-positive noise levels.
detector_footprint_width: Width in pixels for the morphological footprint used in
background subtraction. Larger values remove larger-scale background variations
but may erode colony edges.
detector_noise_radius: Radius in pixels for the morphological opening used to remove
small noise artifacts. Larger values remove larger noise but may erode fine
colony features.
detector_smoothing_sigma: Standard deviation for Gaussian smoothing of intensity
profiles before peak detection. Higher values increase robustness to noise but may
merge nearby peaks. Set to 0 to disable smoothing.
detector_min_peak_distance: Minimum distance between detected peaks in pixels. If None,
automatically estimated from grid dimensions. Prevents detection of spurious peaks
too close together.
detector_peak_prominence: Minimum prominence of peaks for detection. If None,
automatically estimated from signal statistics. Higher values are more selective.
detector_edge_refinement: Whether to refine grid edges using local intensity profiles.
Improves accuracy but adds computational cost.
detector_selection_mode: Strategy for selecting the primary grid when multiple
candidate grids are found. "dominant" selects the grid with the most colonies,
"centered" prefers grids near the image center, "regularized" balances colony
count with spatial regularity.
detector_split_merged: Whether to attempt splitting merged colonies that appear as
single large objects. Improves accuracy for dense plates where colonies grow
together but may over-segment irregular colony morphologies.
grid_aligner_axis: Axis along which to align the grid (0 for rows, 1 for columns).
grid_aligner_mode: Padding mode used when rotating the image for alignment.
mask_opener_footprint: Describes the morphological shape for noise removal or
mask refinement. "auto" lets the system adapt, named shapes ("disk", "square",
"diamond") use a standard footprint, an ndarray specifies a custom structuring
element, and None disables opening.
mask_opener_width: Width of the structuring element when using a named shape for
mask opening. Larger values remove more noise but may erode small colonies.
mask_opener_n_iter: Number of iterations for the morphological opening. More
iterations apply stronger smoothing to the mask.
border_remover_size: Specifies the width of the border region to remove. Larger sizes
eliminate edge artifacts and colonies cropped by image edges but may discard valid
colonies near borders.
small_object_min_size: Specifies the size threshold for considering objects as colonies.
Increasing this parameter reduces false detection of small artifacts but risks
ignoring small colonies.
mask_fill_structure: Binary structuring element for hole filling in masks. Larger or
more connected structures fill bigger holes. None uses the default cross-shaped
element.
mask_fill_origin: Origin offset for the structuring element used in hole filling.
texture_scale: Defines the spatial scale(s) at which texture features are measured.
Larger scales focus on macro-textures; smaller scales enhance granular detail
assessment. Can be a list to compute features at multiple scales simultaneously.
texture_quant_lvl: Number of gray levels for quantizing intensity values in texture
analysis. Higher values capture finer intensity distinctions but increase
computation time and may be sensitive to noise.
texture_enhance: Whether to apply contrast enhancement to the texture input before
computing GLCM features. May improve texture discrimination for low-contrast
colonies.
texture_warn: Boolean that enables warnings when texture measurements may not be
reliable. Use this to flag potential inconsistencies in the captured texture data
or image quality issues.
color_white_chroma_max: Maximum chroma value for classifying a colony as white.
Colonies with chroma below this threshold are labeled as white/achromatic.
color_chroma_min: Minimum chroma value for assigning a chromatic hue. Colonies with
chroma below this value receive a neutral color classification.
color_include_XYZ: Whether to include CIE XYZ color space measurements in addition
to the standard Lab and descriptive color features.
benchmark: Enables time benchmarking for each pipeline step. Useful for performance
debugging but adds overhead to the computation.
verbose: Specifies whether to output detailed process information during execution.
True provides step-by-step logs, which are useful for debugging, while False
ensures silent execution suitable for batch processing.
"""
# Construct the operations pipeline
detector_kwargs = dict(
thresh_method=detector_thresh_method,
subtract_background=detector_subtract_background,
remove_noise=detector_remove_noise,
footprint_width=detector_footprint_width,
noise_radius=detector_noise_radius,
smoothing_sigma=detector_smoothing_sigma,
min_peak_distance=detector_min_peak_distance,
peak_prominence=detector_peak_prominence,
edge_refinement=detector_edge_refinement,
selection_mode=detector_selection_mode,
split_merged=detector_split_merged,
)
ops = [
EnhanceBlockMatch(sigma_psd=bm3d_sigma, block_size=bm3d_block_size,
stage_arg=bm3d_stage_arg, clip=bm3d_clip),
EnhanceLocalContrast(kernel_size=clahe_kernel_size,
clip_limit=clahe_clip_limit),
MedianFilter(mode=median_mode, shape=median_shape, width=median_radius,
cval=median_cval),
# First detection pass
RoundPeaksDetector(**detector_kwargs),
MaskOpening(shape=mask_opener_footprint, width=mask_opener_width,
n_iter=mask_opener_n_iter),
RemoveBorderObjects(border_size=border_remover_size),
SmallObjectRemover(min_size=small_object_min_size),
MaskFill(structure=mask_fill_structure, origin=mask_fill_origin),
GridOversizedObjectRemover(),
ReduceSectionsByLine(),
GridAligner(axis=grid_aligner_axis, mode=grid_aligner_mode),
# Second detection pass
RoundPeaksDetector(**detector_kwargs),
MaskOpening(shape=None),
RemoveBorderObjects(border_size=border_remover_size),
SmallObjectRemover(min_size=small_object_min_size),
GridOversizedObjectRemover(),
MaskFill(structure=mask_fill_structure, origin=mask_fill_origin),
ReduceSectionsByLine(),
]
meas = [
MeasureShape(),
MeasureColor(
include_XYZ=color_include_XYZ,
),
MeasureTexture(
scale=texture_scale,
quant_lvl=texture_quant_lvl,
enhance=texture_enhance,
warn=texture_warn,
),
MeasureIntensity(),
]
super().__init__(ops=ops, meas=meas, benchmark=benchmark, verbose=verbose)
__all__ = ("HeavyRoundPeaksPipeline",)
