Forked from
Swain Lab / aliby / aliby-mirror
7 commits behind, 45 commits ahead of the upstream repository.
extractor.py 27.71 KiB
import typing as t
import bottleneck as bn
import h5py
import numpy as np
import pandas as pd
from agora.abc import ParametersABC, StepABC
from agora.io.cells import Cells
from agora.io.writer import Writer, load_attributes
from aliby.tile.tiler import Tiler
from extraction.core.functions.defaults import exparams_from_meta
from extraction.core.functions.distributors import reduce_z, trap_apply
from extraction.core.functions.loaders import (
load_custom_args,
load_funs,
load_redfuns,
)
# define types
reduction_method = t.Union[t.Callable, str, None]
extraction_tree = t.Dict[
str, t.Dict[reduction_method, t.Dict[str, t.Collection]]
]
extraction_result = t.Dict[
str, t.Dict[reduction_method, t.Dict[str, t.Dict[str, pd.Series]]]
]
# Global variables used to load functions that either analyse cells
# or their background. These global variables both allow the functions
# to be stored in a dictionary for access only on demand and to be
# defined simply in extraction/core/functions.
CELL_FUNS, TRAP_FUNS, ALL_FUNS = load_funs()
CUSTOM_FUNS, CUSTOM_ARGS = load_custom_args()
RED_FUNS = load_redfuns()
class ExtractorParameters(ParametersABC):
"""Base class to define parameters for extraction."""
def __init__(
self,
tree: extraction_tree,
sub_bg: set = set(),
multichannel_ops: t.Dict = {},
):
"""
Initialise.
Parameters
----------
tree: dict
Nested dictionary indicating channels, reduction functions and
metrics to be used.
str channel -> U(function, None) reduction -> str metric
If not of depth three, tree will be filled with None.
sub_bg: set
multichannel_ops: dict
"""
self.tree = tree
self.sub_bg = sub_bg
self.multichannel_ops = multichannel_ops
@classmethod
def default(cls):
return cls({})
@classmethod
def from_meta(cls, meta):
return cls(**exparams_from_meta(meta))
class Extractor(StepABC):
"""
Apply a metric to cells identified in the tiles.
Using the cell masks, the Extractor applies a metric, such as
area or median, to cells identified in the image tiles.
Its methods require both tile images and masks.
Usually the metric is applied to only a tile's masked area, but
some metrics depend on the whole tile.
Extraction follows a three-level tree structure. Channels, such
as GFP, are the root level; the reduction algorithm, such as
maximum projection, is the second level; the specific metric,
or operation, to apply to the masks, such as mean, is the third
or leaf level.
"""
# TODO Alan: Move this to a location with the SwainLab defaults
default_meta = {
"pixel_size": 0.236,
"z_size": 0.6,
"spacing": 0.6,
}
def __init__(
self,
parameters: ExtractorParameters,
store: t.Optional[str] = None,
tiler: t.Optional[Tiler] = None,
):
"""
Initialise Extractor.
Parameters
----------
parameters: core.extractor Parameters
Parameters that include the channels, reduction and
extraction functions.
store: str
Path to the h5 file containing the cell masks.
tiler: pipeline-core.core.segmentation tiler
Class that contains or fetches the images used for
segmentation.
"""
self.params = parameters
if store:
self.local = store
self.load_meta()
else:
# if no h5 file, use the parameters directly
self.meta = {"channel": parameters.to_dict()["tree"].keys()}
if tiler:
self.tiler = tiler
available_channels = set((*tiler.channels, "general"))
# only extract for channels available
self.params.tree = {
k: v
for k, v in self.params.tree.items()
if k in available_channels
}
self.params.sub_bg = available_channels.intersection(
self.params.sub_bg
)
# add background subtracted channels to those available
available_channels_bgsub = available_channels.union(
[c + "_bgsub" for c in self.params.sub_bg]
)
# remove any multichannel operations requiring a missing channel
for op, (input_ch, _, _) in self.params.multichannel_ops.items():
if not set(input_ch).issubset(available_channels_bgsub):
self.params.multichannel_ops.pop(op)
self.load_funs()
@classmethod
def from_tiler(
cls,
parameters: ExtractorParameters,
store: str,
tiler: Tiler,
):
"""Initiate from a tiler instance."""
return cls(parameters, store=store, tiler=tiler)
@classmethod
def from_img(
cls,
parameters: ExtractorParameters,
store: str,
img_meta: tuple,
):
"""Initiate from images."""
return cls(parameters, store=store, tiler=Tiler(*img_meta))
@property
def channels(self):
"""Get a tuple of the available channels."""
if not hasattr(self, "_channels"):
if type(self.params.tree) is dict:
self._channels = tuple(self.params.tree.keys())
return self._channels
@property
def current_position(self):
return str(self.local).split("/")[-1][:-3]
@property
def group(self):
"""Return path within the h5 file."""
if not hasattr(self, "_out_path"):
self._group = "/extraction/"
return self._group
def load_custom_funs(self):
"""
Incorporate the extra arguments of custom functions into their
definitions.
Normal functions only have cell_masks and trap_image as their
arguments, and here custom functions are made the same by
setting the values of their extra arguments.
Any other parameters are taken from the experiment's metadata
and automatically applied. These parameters therefore must be
loaded within an Extractor instance.
"""
# find functions specified in params.tree
funs = set(
[
fun
for ch in self.params.tree.values()
for red in ch.values()
for fun in red
]
)
# consider only those already loaded from CUSTOM_FUNS
funs = funs.intersection(CUSTOM_FUNS.keys())
# find their arguments
self._custom_arg_vals = {
k: {k2: self.get_meta(k2) for k2 in v}
for k, v in CUSTOM_ARGS.items()
}
# define custom functions
self._custom_funs = {}
for k, f in CUSTOM_FUNS.items():
def tmp(f):
# pass extra arguments to custom function
# return a function of cell_masks and trap_image
return lambda cell_masks, trap_image: trap_apply(
f,
cell_masks,
trap_image,
**self._custom_arg_vals.get(k, {}),
)
self._custom_funs[k] = tmp(f)
def load_funs(self):
"""Define all functions, including custom ones."""
self.load_custom_funs()
self._all_cell_funs = set(self._custom_funs.keys()).union(CELL_FUNS)
# merge the two dicts
self._all_funs = {**self._custom_funs, **ALL_FUNS}
def load_meta(self):
"""Load metadata from h5 file."""
self.meta = load_attributes(self.local)
def get_tiles(
self,
tp: int,
channels: t.Optional[t.List[t.Union[str, int]]] = None,
z: t.Optional[t.List[str]] = None,
**kwargs,
) -> t.Optional[np.ndarray]:
"""
Find tiles for a given time point, channels, and z-stacks.
Any additional keyword arguments are passed to
tiler.get_tiles_timepoint
Parameters
----------
tp: int
Time point of interest.
channels: list of strings (optional)
Channels of interest.
z: list of integers (optional)
Indices for the z-stacks of interest.
"""
if channels is None:
# find channels from tiler
channel_ids = list(range(len(self.tiler.channels)))
elif len(channels):
# a subset of channels was specified
channel_ids = [self.tiler.get_channel_index(ch) for ch in channels]
else:
# a list of the indices of the z stacks
channel_ids = None
if z is None:
# include all Z channels
z = list(range(self.tiler.shape[-3]))
# get the image data via tiler
res = (
self.tiler.get_tiles_timepoint(
tp, channels=channel_ids, z=z, **kwargs
)
if channel_ids
else None
)
# res has dimensions (tiles, channels, 1, Z, X, Y)
return res
def extract_traps(
self,
traps: t.List[np.ndarray],
masks: t.List[np.ndarray],
cell_property: str,
cell_labels: t.Dict[int, t.List[int]],
) -> t.Tuple[t.Union[t.Tuple[float], t.Tuple[t.Tuple[int]]]]:
"""
Apply a function to a whole position.
Parameters
----------
traps: list of arrays
t.List of images.
masks: list of arrays
t.List of masks.
cell_property: str
Property to extract, including imBackground.
cell_labels: dict
A dict of cell labels with trap_ids as keys and a list
of cell labels as values.
Returns
-------
res_idx: a tuple of tuples
A two-tuple comprising a tuple of results and a tuple of
the tile_id and cell labels
"""
if cell_labels is None:
self._log("No cell labels given. Sorting cells using index.")
cell_fun = True if cell_property in self._all_cell_funs else False
idx = []
results = []
for trap_id, (mask_set, trap, local_cell_labels) in enumerate(
zip(masks, traps, cell_labels.values())
):
# ignore empty traps
if len(mask_set):
# find property from the tile
result = self._all_funs[cell_property](mask_set, trap)
if cell_fun:
# store results for each cell separately
for cell_label, val in zip(local_cell_labels, result):
results.append(val)
idx.append((trap_id, cell_label))
else:
# background (trap) function
results.append(result)
idx.append(trap_id)
res_idx = (tuple(results), tuple(idx))
return res_idx
def extract_funs(
self,
traps: t.List[np.array],
masks: t.List[np.array],
cell_properties: t.List[str],
**kwargs,
) -> t.Dict[str, pd.Series]:
"""
Return dict with metrics as key and cell_properties as values.
Data from one time point is used.
"""
d = {
cell_property: self.extract_traps(
traps=traps, masks=masks, cell_property=cell_property, **kwargs
)
for cell_property in cell_properties
}
return d
def reduce_extract(
self,
traps: np.ndarray,
masks: t.List[np.ndarray],
tree_branch: t.Dict[reduction_method, t.Collection[str]],
**kwargs,
) -> t.Dict[str, t.Dict[reduction_method, t.Dict[str, pd.Series]]]:
"""
Wrapper to reduce to a 2D image and then extract.
Parameters
----------
tiles_data: array
An array of image data arranged as (tiles, X, Y, Z)
masks: list of arrays
An array of masks for each trap: one per cell at the trap
tree_branch: dict
An upper branch of the extraction tree: a dict for which
keys are reduction functions and values are either a list
or a set of strings giving the cell properties to be found.
For example: {'np_max': {'max5px', 'mean', 'median'}}
**kwargs: dict
All other arguments passed to Extractor.extract_funs.
Returns
------
Dict of dataframes with the corresponding reductions and metrics nested.
"""
# FIXME hack to pass tests
if "labels" in kwargs:
kwargs["cell_labels"] = kwargs.pop("labels")
# create dict with keys naming the reduction in the z-direction
# and the reduced data as values
reduced_tiles = {}
if traps is not None:
for red_fun in tree_branch.keys():
reduced_tiles[red_fun] = [
self.reduce_dims(tile_data, method=RED_FUNS[red_fun])
for tile_data in traps
]
# calculate cell and tile properties
d = {
red_fun: self.extract_funs(
cell_properties=cell_properties,
traps=reduced_tiles.get(red_fun, [None for _ in masks]),
masks=masks,
**kwargs,
)
for red_fun, cell_properties in tree_branch.items()
}
return d
def reduce_dims(
self, img: np.ndarray, method: reduction_method = None
) -> np.ndarray:
"""
Collapse a z-stack into 2d array using method.
If method is None, return the original data.
Parameters
----------
img: array
An array of the image data arranged as (X, Y, Z).
method: function
The reduction function.
"""
reduced = img
if method is not None:
reduced = reduce_z(img, method)
return reduced
def make_tree_bits(self, tree):
"""Put extraction tree and information for the channels into a dict."""
if tree is None:
# use default
tree: extraction_tree = self.params.tree
tree_bits = {
"tree": tree,
# dictionary with channel: {reduction algorithm : metric}
"channel_tree": {
ch: v for ch, v in tree.items() if ch != "general"
},
}
# tuple of the fluorescence channels
tree_bits["tree_channels"] = (*tree_bits["channel_tree"],)
return tree_bits
def get_masks(self, tp, masks, cells):
"""Get the masks as a list with an array of masks for each trap."""
# find the cell masks for a given trap as a dict with trap_ids as keys
if masks is None:
raw_masks = cells.at_time(tp, kind="mask")
masks = {trap_id: [] for trap_id in range(cells.ntraps)}
for trap_id, cells in raw_masks.items():
if len(cells):
masks[trap_id] = np.stack(np.array(cells)).astype(bool)
# convert to a list of masks
# one array of size (no cells, tile_size, tile_size) per trap
masks = [np.array(v) for v in masks.values()]
return masks
def get_cell_labels(self, tp, cell_labels, cells):
"""Get the cell labels per trap as a dict with trap_ids as keys."""
if cell_labels is None:
raw_cell_labels = cells.labels_at_time(tp)
cell_labels = {
trap_id: raw_cell_labels.get(trap_id, [])
for trap_id in range(cells.ntraps)
}
return cell_labels
def get_background_masks(self, masks, tile_size):
"""
Generate boolean background masks.
Combine masks per trap and then take the logical inverse.
"""
if self.params.sub_bg:
bgs = ~np.array(
list(
map(
# sum over masks for each cell
lambda x: np.sum(x, axis=0)
if np.any(x)
else np.zeros((tile_size, tile_size)),
masks,
)
)
).astype(bool)
else:
bgs = np.array([])
return bgs
def extract_one_channel(
self, tree_bits, cell_labels, tiles, masks, bgs, **kwargs
):
"""
Extract using all metrics requiring a single channel.
Apply first without and then with background subtraction.
Return the extraction results and a dict of background
corrected images.
"""
d = {}
img_bgsub = {}
for ch, tree_branch in tree_bits["tree"].items():
# NB ch != is necessary for threading
if ch != "general" and tiles is not None and len(tiles):
# image data for all traps for a particular channel and time point
# arranged as (traps, Z, X, Y)
# we use 0 here to access the single time point available
img = tiles[:, tree_bits["tree_channels"].index(ch), 0]
else:
# no reduction applied to bright-field images
img = None
# apply metrics to image data
d[ch] = self.reduce_extract(
traps=img,
masks=masks,
tree_branch=tree_branch,
cell_labels=cell_labels,
**kwargs,
)
# apply metrics to image data with the background subtracted
if bgs.any() and ch in self.params.sub_bg and img is not None:
# calculate metrics with background subtracted
ch_bs = ch + "_bgsub"
# subtract median background
bgsub_mapping = map(
# move Z to last column to allow subtraction
lambda img, bgs: np.moveaxis(img, 0, -1)
# median of background over all pixels for each Z section
- bn.median(img[:, bgs], axis=1),
img,
bgs,
)
# apply map and convert to array
mapping_result = np.stack(list(bgsub_mapping))
# move Z axis back to the second column
img_bgsub[ch_bs] = np.moveaxis(mapping_result, -1, 1)
# apply metrics to background-corrected data
d[ch_bs] = self.reduce_extract(
tree_branch=tree_bits["channel_tree"][ch],
traps=img_bgsub[ch_bs],
masks=masks,
cell_labels=cell_labels,
**kwargs,
)
return d, img_bgsub
def extract_multiple_channels(
self, tree_bits, cell_labels, tiles, masks, **kwargs
):
"""
Extract using all metrics requiring multiple channels.
"""
available_chs = set(self.img_bgsub.keys()).union(
tree_bits["tree_channels"]
)
d = {}
for name, (
chs,
reduction_fun,
op,
) in self.params.multichannel_ops.items():
common_chs = set(chs).intersection(available_chs)
# all required channels should be available
if len(common_chs) == len(chs):
channels_stack = np.stack(
[
self.get_imgs(ch, tiles, tree_bits["tree_channels"])
for ch in chs
],
axis=-1,
)
# reduce in Z
traps = RED_FUNS[reduction_fun](channels_stack, axis=1)
# evaluate multichannel op
if name not in d:
d[name] = {}
if reduction_fun not in d[name]:
d[name][reduction_fun] = {}
d[name][reduction_fun][op] = self.extract_traps(
traps,
masks,
op,
cell_labels,
)
return d
def extract_tp(
self,
tp: int,
tree: t.Optional[extraction_tree] = None,
tile_size: int = 117,
masks: t.Optional[t.List[np.ndarray]] = None,
cell_labels: t.Optional[t.List[int]] = None,
**kwargs,
) -> t.Dict[str, t.Dict[str, t.Dict[str, tuple]]]:
"""
Extract for an individual time point.
Parameters
----------
tp : int
Time point being analysed.
tree : dict
Nested dictionary indicating channels, reduction functions
and metrics to be used.
For example: {'general': {'None': ['area', 'volume', 'eccentricity']}}
tile_size : int
Size of the tile to be extracted.
masks : list of arrays
A list of masks per trap with each mask having dimensions
(ncells, tile_size, tile_size) and with one mask per cell.
cell_labels : dict
A dictionary with trap_ids as keys and cell_labels as values.
**kwargs : keyword arguments
Passed to extractor.reduce_extract.
Returns
-------
d: dict
Dictionary of the results with three levels of dictionaries.
The first level has channels as keys.
The second level has reduction metrics as keys.
The third level has cell or background metrics as keys and a
two-tuple as values.
The first tuple is the result of applying the metrics to a
particular cell or trap; the second tuple is either
(trap_id, cell_label) for a metric applied to a cell or a
trap_id for a metric applied to a trap.
An example is d["GFP"]["np_max"]["mean"][0], which gives a tuple
of the calculated mean GFP fluorescence for all cells.
"""
# dict of information from extraction tree
tree_bits = self.make_tree_bits(tree)
# create a Cells object to extract information from the h5 file
cells = Cells(self.local)
# find the cell labels as dict with trap_ids as keys
cell_labels = self.get_cell_labels(tp, cell_labels, cells)
# get masks one per cell per trap
masks = self.get_masks(tp, masks, cells)
# find image data at the time point
# stored as an array arranged as (traps, channels, 1, Z, X, Y)
tiles = self.get_tiles(
tp, tile_shape=tile_size, channels=tree_bits["tree_channels"]
)
# generate boolean masks for background for each trap
bgs = self.get_background_masks(masks, tile_size)
# perform extraction
res_one, self.img_bgsub = self.extract_one_channel(
tree_bits, cell_labels, tiles, masks, bgs, **kwargs
)
res_multiple = self.extract_multiple_channels(
tree_bits, cell_labels, tiles, masks, **kwargs
)
res = {**res_one, **res_multiple}
return res
def get_imgs(self, channel: t.Optional[str], tiles, channels=None):
"""
Return image from a correct source, either raw or bgsub.
Parameters
----------
channel: str
Name of channel to get.
tiles: ndarray
An array of the image data having dimensions of
(tile_id, channel, tp, tile_size, tile_size, n_zstacks).
channels: list of str (optional)
t.List of available channels.
Returns
-------
img: ndarray
An array of image data with dimensions
(no tiles, X, Y, no Z channels)
"""
if channels is None:
channels = (*self.params.tree,)
if channel in channels: # TODO start here to fetch channel using regex
return tiles[:, channels.index(channel), 0]
elif channel in self.img_bgsub:
return self.img_bgsub[channel]
def _run_tp(
self,
tps: t.List[int] = None,
tree=None,
save=True,
**kwargs,
) -> dict:
"""
Run extraction for one position and for the specified time points.
Save the results to a h5 file.
Parameters
----------
tps: list of int (optional)
Time points to include.
tree: dict (optional)
Nested dictionary indicating channels, reduction functions and
metrics to be used.
For example: {'general': {'None': ['area', 'volume', 'eccentricity']}}
save: boolean (optional)
If True, save results to h5 file.
kwargs: keyword arguments (optional)
Passed to extract_tp.
Returns
-------
d: dict
A dict of the extracted data for one position with a concatenated
string of channel, reduction metric, and cell metric as keys and
pd.DataFrame of the extracted data for all time points as values.
"""
if tree is None:
tree = self.params.tree
if tps is None:
tps = list(range(self.meta["time_settings/ntimepoints"][0]))
elif isinstance(tps, int):
tps = [tps]
# store results in dict
d = {}
for tp in tps:
# extract for each time point and convert to dict of pd.Series
new = flatten_nesteddict(
self.extract_tp(tp=tp, tree=tree, **kwargs),
to="series",
tp=tp,
)
# concatenate with data extracted from early time points
for k in new.keys():
d[k] = pd.concat((d.get(k, None), new[k]), axis=1)
# add indices to pd.Series containing the extracted data
for k in d.keys():
indices = ["experiment", "position", "trap", "cell_label"]
idx = (
indices[-d[k].index.nlevels :]
if d[k].index.nlevels > 1
else [indices[-2]]
)
d[k].index.names = idx
# save
if save:
self.save_to_h5(d)
return d
def save_to_h5(self, dict_series, path=None):
"""
Save the extracted data for one position to the h5 file.
Parameters
----------
dict_series: dict
A dictionary of the extracted data, created by run.
path: Path (optional)
To the h5 file.
"""
if path is None:
path = self.local
self.writer = Writer(path)
for extract_name, series in dict_series.items():
dset_path = "/extraction/" + extract_name
self.writer.write(dset_path, series)
self.writer.id_cache.clear()
def get_meta(self, flds: t.Union[str, t.Collection]):
"""Obtain metadata for one or multiple fields."""
if isinstance(flds, str):
flds = [flds]
meta_short = {k.split("/")[-1]: v for k, v in self.meta.items()}
return {
f: meta_short.get(f, self.default_meta.get(f, None)) for f in flds
}
### Helpers
def flatten_nesteddict(
nest: dict, to="series", tp: int = None
) -> t.Dict[str, pd.Series]:
"""
Convert a nested extraction dict into a dict of pd.Series.
Parameters
----------
nest: dict of dicts
Contains the nested results of extraction.
to: str (optional)
Specifies the format of the output, either pd.Series (default)
or a list
tp: int
Time point used to name the pd.Series
Returns
-------
d: dict
A dict with a concatenated string of channel, reduction metric,
and cell metric as keys and either a pd.Series or a list of the
corresponding extracted data as values.
"""
d = {}
for k0, v0 in nest.items():
for k1, v1 in v0.items():
for k2, v2 in v1.items():
d["/".join((k0, k1, k2))] = (
pd.Series(*v2, name=tp) if to == "series" else v2
)
return d