from __future__ import annotations
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from phenotypic._core._grid_image import GridImage
from phenotypic.abc_ import ObjectRefiner
from phenotypic.abc_ import GridOperation
from phenotypic.sdk_.exceptions_ import GridImageInputError
from phenotypic.sdk_.funcs_ import validate_operation_integrity
[docs]
class GridObjectRefiner(ObjectRefiner, GridOperation, ABC):
"""Abstract base class for post-detection refinement operations on grid-aligned plate images.
GridObjectRefiner is the grid-aware variant of ObjectRefiner, combining object mask refinement
with grid structure awareness. It refines detected objects (colony masks and labeled maps) while
respecting well boundaries and grid-aligned regions in arrayed plate images (96-well, 384-well,
etc.). Like ObjectRefiner, it protects original image data (RGB, grayscale, detection matrix)
and modifies only detection results.
**Quick Decision Guide: GridObjectRefiner vs ObjectRefiner**
- **Use GridObjectRefiner if:** Refining detections on GridImage where well structure or grid
position affects refinement logic (per-well filtering, boundary enforcement, position-aware cleanup).
- **Use ObjectRefiner if:** Refining on plain Image without grid, or using global algorithms that
don't need grid awareness (size filtering, shape filtering, general morphology).
- **GridImage requirement:** GridObjectRefiner only accepts GridImage input; plain Image raises
GridImageInputError at runtime.
- **Grid-aware refinement:** Access well positions, row/column boundaries, and grid metadata via
image.grid to make position-aware decisions (e.g., filter colonies oversized for their well).
- **Oversized colonies:** [GridOversizedObjectRemover](src/phenotypic/refine/_grid_oversized_object_remover.py)
removes objects spanning nearly entire well (merged colonies, segmentation spillover).
- **Per-well largest:** [KeepSectionLargest](src/phenotypic/refine/_keep_section_largest.py) keeps
only the largest object per grid cell (one colony per well).
- **Multi-well reducer:** [ReduceSectionsByLine](src/phenotypic/refine/_reduce_multiple_grid_objects.py)
merges multiple objects per well into single representative region.
- **Border handling:** Grid structure enables identifying and filtering objects near plate/well
edges that may be incomplete or distorted.
- **Grid registration:** When grid detection is imperfect, grid-aware refinement helps filter
mis-assigned or boundary-adjacent objects by position.
- **When to chain:** Combine grid refiners in ImagePipeline (e.g., global size filter, then
per-well filtering) for comprehensive cleanup respecting array structure.
**What is GridObjectRefiner?**
GridObjectRefiner is the specialized version of ObjectRefiner for GridImage objects, adding grid-aware
refinement to the core post-detection cleanup workflow:
- **Grid-structure awareness:** Unlike ObjectRefiner (which operates globally), GridObjectRefiner can
access grid boundaries, well positions, row/column indices, and per-cell metadata via ``image.grid``.
This enables refinement logic that respects array structure.
- **GridImage requirement:** Accepts only GridImage input (with detected grid structure), enforced
at runtime via ``GridImageInputError``. Passing a plain Image raises an error.
- **Grid access interface:** Within ``_operate()``, call ``image.grid.get_row_edges()``,
``image.grid.get_col_edges()``, and ``image.grid.info()`` to retrieve grid metadata and make
position-aware refinement decisions.
- **Detection-only modification:** Like ObjectRefiner, modifies only ``image.objmask[:]`` and
``image.objmap[:]``. Original image components (RGB, grayscale, detection matrix) are protected
via ``@validate_operation_integrity`` decorator.
- **Array phenotyping:** Well-suited for high-throughput plate analysis where grid structure matters
for biology (one colony per well expected, oversized objects indicate merging, boundary objects
may be incomplete or distorted).
**When to use GridObjectRefiner vs ObjectRefiner**
- **ObjectRefiner:** Use when refining detections on a plain Image without grid structure.
Examples: general-purpose size filtering, morphological cleanup, shape filtering (applies
globally regardless of position).
- **GridObjectRefiner:** Use when refining detections on a GridImage where well structure matters.
Examples: removing objects larger than their grid cell (``GridOversizedObjectRemover``),
per-well filtering, grid-aligned edge removal. The grid structure enables position-aware
refinement that improves array phenotyping accuracy.
**Typical Use Cases**
GridObjectRefiner is useful for addressing grid-specific artifacts:
- **Oversized colonies:** Objects spanning nearly an entire well (merged colonies, agar edges,
segmentation spillover). Filtering improves per-well consistency.
- **Inter-well artifacts:** Detections touching or bridging grid cell boundaries from uneven
lighting or thresholding errors.
- **Boundary contamination:** Colonies near plate edges that are incomplete or distorted.
Grid structure allows identifying and filtering boundary-adjacent objects.
- **Grid registration errors:** When grid detection is imperfect, some objects may be mis-assigned
to wells; grid-aware refinement can filter or relocate based on position.
**Implementing a Custom GridObjectRefiner**
Subclass GridObjectRefiner and implement ``_operate()``. Use this template:
.. code-block:: python
from phenotypic.abc_ import GridObjectRefiner
from phenotypic import GridImage
import numpy as np
class MyGridRefiner(GridObjectRefiner):
max_width_fraction: float = 0.9 # Annotated class-level field
def _operate(self, image: GridImage) -> GridImage:
# Get grid structure information
col_edges = image.grid.get_col_edges() # x-coordinates of column boundaries
row_edges = image.grid.get_row_edges() # y-coordinates of row boundaries
nrows, ncols = image.grid.nrows, image.grid.ncols
# Get per-object grid metadata (label, row, column, position)
grid_info = image.grid.info() # pd.DataFrame
# Measure object properties
objmap = image.objmap[:]
from skimage.measure import regionprops_table
props = regionprops_table(objmap, properties=['label', 'bbox', 'area'])
# Make position-aware refinement decisions
# Example: filter objects by grid position or size relative to well
max_width = (col_edges[1:] - col_edges[:-1]).max()
# ... implement your grid-aware logic ...
return image
**Key Rules**
1. ``_operate()`` must be an instance method (access parameters via ``self``).
2. All parameters except ``image`` must be declared as annotated class-level fields.
3. Only modify ``image.objmask[:]`` and ``image.objmap[:]``.
4. Access grid via ``image.grid`` methods: get_row_edges(), get_col_edges(), info().
5. Return the modified GridImage object.
**Per-Well Filtering Patterns**
Common grid-aware refinement strategies:
- **Remove oversized objects:** Filter objects larger than well dimensions (merged colonies, segmentation
spillover). Compare object bounding box to max_width and max_height of grid cells.
- **Keep largest per well:** Select only the largest object per grid cell (assumes one colony per well).
Use grid_info to group objects by row and column, then keep max-area object per group.
- **Remove boundary objects:** Filter objects touching or near well edges (incomplete detections).
Use grid_info's boundary flags or compute distance to nearest grid boundary.
- **Per-row/column filtering:** Apply different thresholds by row or column position (accounts for
uneven illumination across plate). Use grid_info to stratify objects by position.
**Grid Access Interface**
Within ``_operate()``, use the GridImage accessor to retrieve grid metadata:
.. code-block:: python
# Grid structure information
nrows, ncols = image.grid.nrows, image.grid.ncols
row_edges = image.grid.get_row_edges() # Row boundary positions (y-coordinates)
col_edges = image.grid.get_col_edges() # Col boundary positions (x-coordinates)
grid_info = image.grid.info() # DataFrame with per-object grid metadata
# grid_info columns (example):
# ['label', 'row', 'col', 'row_edge_min', 'row_edge_max', 'col_edge_min', 'col_edge_max', ...]
# Per-well cell dimensions
cell_heights = row_edges[1:] - row_edges[:-1]
cell_widths = col_edges[1:] - col_edges[:-1]
Notes:
- **GridImage input required:** ``apply()`` enforces GridImage type at runtime.
Passing a plain Image raises ``GridImageInputError``.
- **Protected components:** The ``@validate_operation_integrity`` decorator ensures
``image.rgb``, ``image.gray``, ``image.detect_mat`` cannot be modified.
Only ``image.objmask`` and ``image.objmap`` can be refined.
- **Immutability by default:** ``apply(image)`` returns a modified copy. Set
``inplace=True`` for memory-efficient in-place modification.
- **Grid structure assumption:** Your algorithm should assume a valid, registered grid.
If grid metadata is unreliable, refinement may fail or produce wrong results.
- **Instance _operate() method:** ``_operate()`` is an instance method; access parameters via ``self``.
- **Field-based parameters:** All ``_operate()`` parameters except ``image``
are declared as annotated class-level fields.
Examples:
Remove objects larger than their grid cell width:
>>> from phenotypic.abc_ import GridObjectRefiner
>>> from phenotypic import GridImage
>>> import numpy as np
>>> class OversizedObjectRemover(GridObjectRefiner):
... '''Remove objects exceeding cell dimensions.'''
...
... def _operate(self, image: GridImage) -> GridImage:
... # Get grid boundaries
... col_edges = image.grid.get_col_edges()
... row_edges = image.grid.get_row_edges()
... max_width = (col_edges[1:] - col_edges[:-1]).max()
... max_height = (row_edges[1:] - row_edges[:-1]).max()
... # Measure objects
... objmap = image.objmap[:]
... from skimage.measure import regionprops_table
... props = regionprops_table(objmap, properties=['label', 'bbox'])
... # Filter oversized
... import pandas as pd
... df = pd.DataFrame(props)
... df['width'] = df['bbox-2'] - df['bbox-0']
... df['height'] = df['bbox-3'] - df['bbox-1']
... keep = df[(df['width'] < max_width) &
... (df['height'] < max_height)]['label'].values
... # Refine map
... refined = np.where(np.isin(objmap, keep), objmap, 0)
... image.objmap[:] = refined
... return image
>>> # Usage on gridded plate image
>>> from phenotypic.detect import OtsuDetector
>>> image = GridImage.imread('plate.jpg', nrows=8, ncols=12)
>>> detected = OtsuDetector().apply(image)
>>> cleaned = OversizedObjectRemover().apply(detected)
Per-well filtering: remove colonies oversized for their well:
>>> from phenotypic.abc_ import GridObjectRefiner
>>> from phenotypic import GridImage
>>> from phenotypic.data import load_synth_yeast_plate
>>> import numpy as np
>>> import pandas as pd
>>> class PerWellOversizedRemover(GridObjectRefiner):
... '''Remove objects that exceed size threshold relative to their well.'''
...
... max_area_fraction: float = 0.8
...
... def _operate(self, image: GridImage) -> GridImage:
... # Get grid structure
... col_edges = image.grid.get_col_edges()
... row_edges = image.grid.get_row_edges()
... # Compute well area (assume uniform grid cells)
... cell_width = (col_edges[1:] - col_edges[:-1]).mean()
... cell_height = (row_edges[1:] - row_edges[:-1]).mean()
... max_cell_area = cell_width * cell_height
... # Get grid info and measure object areas
... objmap = image.objmap[:]
... grid_info = image.grid.info()
... from skimage.measure import regionprops_table
... props = regionprops_table(objmap, properties=['label', 'area'])
... props_df = pd.DataFrame(props)
... # Filter: keep objects smaller than max_area_fraction of their well
... max_allowed_area = max_cell_area * max_area_fraction
... keep = props_df[props_df['area'] < max_allowed_area]['label'].values
... # Refine map
... refined = np.where(np.isin(objmap, keep), objmap, 0)
... image.objmap[:] = refined
... return image
>>> # Usage: remove merged colonies and artifacts spanning most of well
>>> from phenotypic.detect import OtsuDetector
>>> image = load_synth_yeast_plate() # Returns GridImage
>>> detected = OtsuDetector().apply(image)
>>> # Remove colonies > 80% of well area (likely merged or segmentation error)
>>> cleaner = PerWellOversizedRemover(max_area_fraction=0.8)
>>> refined = cleaner.apply(detected)
>>> print(f"Removed oversized: {detected.objmap[:].max()} -> {refined.objmap[:].max()}")
Chaining grid and non-grid refinements:
>>> from phenotypic import GridImage, ImagePipeline
>>> from phenotypic.detect import OtsuDetector
>>> from phenotypic.refine import SmallObjectRemover, GridOversizedObjectRemover
>>> # Create detection pipeline with mixed refinements
>>> pipeline = ImagePipeline()
>>> pipeline.add(OtsuDetector()) # Detect colonies
>>> pipeline.add(SmallObjectRemover(min_size=100)) # Global size filter
>>> pipeline.add(GridOversizedObjectRemover()) # Grid-aware filter
>>> # Apply to gridded plate
>>> image = GridImage.imread('plate.jpg', nrows=8, ncols=12)
>>> results = pipeline.operate([image])
>>> refined_image = results[0]
>>> print(f"Refined: {refined_image.objmap[:].max()} colonies")
"""
[docs]
@validate_operation_integrity("image.rgb", "image.gray", "image.detect_mat")
def apply(self, image: GridImage, inplace: bool = False) -> GridImage:
from phenotypic import GridImage
if not isinstance(image, GridImage):
raise GridImageInputError()
output = super().apply(image=image, inplace=inplace)
return output
@abstractmethod
def _operate(self, image: GridImage) -> GridImage: # type: ignore[override]
return image
