aliby.track.cell_tracker.CellTracker

class CellTracker(feats2use=None, trapfeats=None, extrafeats=None, model=None, bak_model=None, thresh=None, low_thresh=None, high_thresh=None, nstepsback=None, aweights=None, red_fun=None, max_distance=None, **kwargs)

Bases: FeatureCalculator

Class used to manage cell tracking. You can call it using an existing model or use the inherited CellTrainer to get a new one.

Initialization parameters:

Model

sklearn.ensemble.RandomForestClassifier object

Trapfeats

Features to manually calculate within a trap

Extrafeats

Additional features to calculate

Model

Model to use, if provided ignores all other args but threshs

Bak_model

Backup mode to use when prediction is unsure

Nstepsback

int Number of timepoints to go back

Thresh

float Cut-off value to assume a cell is not new

Low_thresh

Lower thresh for model switching

High_thresh

Higher thresh for model switching. Probabilities between these two thresholds summon the backup model.

Aweights

area weight for barycentre calculations

# Feature order in array features 1. basic features 2. trap features (within a trap) 3. extra features (process between imgs)

Attributes

a_ind

ma_ind

x_ind

y_ind

Methods

assign_lbls(prob_backtrace, prev_lbls[, red_fun])	Assign labels using a prediction matrix of nxmxl where n is the number of cells in the previous image, m the number of steps back considered and l in the new image.
calc_dtfeats(n3darray)	Calculates features obtained between timepoints, such as distance for every pair of cells from t1 to t2.
calc_feat_ndarray(prev_feats, new_feats)	Calculate feature ndarray using two ndarrays of features.
calc_feats_from_mask(masks[, feats2use, ...])	Calculate feature ndarray from ndarray of cell masks
calc_trapfeats(basefeats)	Calculate trap-based features using basic ones. :basefeats: (n basic outfeats) 1-D array with features outputed by skimage.measure.regionprops_table.
get_feats2use()	Return feats to be used from a loaded random forest model model
get_new_lbls(new_img, prev_lbls, prev_feats, ...)	Core function to calculate the new cell labels.
predict_proba_from_ndarray(array_3d[, ...])	input
probabilities_from_impair(image_t1, image_t2)	Convenience function to test tracking between two time-points
run_tp(masks[, state])	Assign labels to new masks using a state dictionary.
scale_feats(feats, pixel_size)	input

get_outfeats
load_model
set_named_ids

__init__(feats2use=None, trapfeats=None, extrafeats=None, model=None, bak_model=None, thresh=None, low_thresh=None, high_thresh=None, nstepsback=None, aweights=None, red_fun=None, max_distance=None, **kwargs)

Methods

__init__([feats2use, trapfeats, extrafeats, ...])
assign_lbls(prob_backtrace, prev_lbls[, red_fun])	Assign labels using a prediction matrix of nxmxl where n is the number of cells in the previous image, m the number of steps back considered and l in the new image.
calc_dtfeats(n3darray)	Calculates features obtained between timepoints, such as distance for every pair of cells from t1 to t2.
calc_feat_ndarray(prev_feats, new_feats)	Calculate feature ndarray using two ndarrays of features.
calc_feats_from_mask(masks[, feats2use, ...])	Calculate feature ndarray from ndarray of cell masks
calc_trapfeats(basefeats)	Calculate trap-based features using basic ones. :basefeats: (n basic outfeats) 1-D array with features outputed by skimage.measure.regionprops_table.
get_feats2use()	Return feats to be used from a loaded random forest model model
get_new_lbls(new_img, prev_lbls, prev_feats, ...)	Core function to calculate the new cell labels.
get_outfeats([feats2use])	rtype tuple
load_model(path, fname)
predict_proba_from_ndarray(array_3d[, ...])	input
probabilities_from_impair(image_t1, image_t2)	Convenience function to test tracking between two time-points
run_tp(masks[, state])	Assign labels to new masks using a state dictionary.
scale_feats(feats, pixel_size)	input
set_named_ids()

Attributes

a_ind
ma_ind
x_ind
y_ind

assign_lbls(prob_backtrace, prev_lbls, red_fun=None)

Assign labels using a prediction matrix of nxmxl where n is the number of cells in the previous image, m the number of steps back considered and l in the new image. It assigns the number zero if it doesn’t find the cell. — input

Prob_backtrace

Probability n x m x l array obtained as an output of rforest

Prev_labels

List of cell labels for previous timepoint to be compared.

Red_fun

Function used to collapse the previous timepoints into one. If none provided it uses maximum and ignores np.nans.

returns

New_lbls

ndarray of newly assigned labels obtained, new cells as

zero.

calc_dtfeats(n3darray)

Calculates features obtained between timepoints, such as distance for every pair of cells from t1 to t2. —

input

N3darray

ndarray (ncells_prev, ncells_new, ntfeats) containing a cell-wise substraction of the features in the input ndarrays.

returns

Newarray

3d array taking the features specified in self.outfeats and self.trapfeats

and adding dtfeats

calc_feat_ndarray(prev_feats, new_feats)

Calculate feature ndarray using two ndarrays of features. —

input

Prev_feats

ndarray (ncells, nfeats) of timepoint 1

New_feats

ndarray (ncells, nfeats) of timepoint 2

returns

N3darray

ndarray (ncells_prev, ncells_new, nfeats) containing a cell-wise substraction of the features in the input ndarrays.

calc_feats_from_mask(masks, feats2use=None, trapfeats=None, scale=True, pixel_size=None)

Calculate feature ndarray from ndarray of cell masks — input

Masks

ndarray (ncells, x_size, y_size), typically dtype bool

Feats2use

list of strings with the feature properties to extract. If it is None it uses the ones set in self.feats2use.

Trapfeats

List of str with additional features to use calculated immediately after basic features.

Scale

bool, if True scales mask to a defined pixel_size.

Pixel_size

float, used to rescale the object features.

returns

(ncells, nfeats) ndarray of features for input masks

calc_trapfeats(basefeats)

Calculate trap-based features using basic ones. :basefeats: (n basic outfeats) 1-D array with features outputed by

skimage.measure.regionprops_table

requires

self.aind self.aweights self.xind self.yind self.trapfeats

returns (ntrapfeats) 1-D array with

get_feats2use()

Return feats to be used from a loaded random forest model model

get_new_lbls(new_img, prev_lbls, prev_feats, max_lbl, new_feats=None, pixel_size=None, **kwargs)

Core function to calculate the new cell labels.

---

input

New_img

ndarray (len, width, ncells) containing the cell outlines

Max_lbl

int indicating the last assigned cell label

Prev_feats

list of ndarrays of size (ncells x nfeatures)

containing the features of previous timepoints :prev_lbls: list of list of ints corresponding to the cell labels in

the previous timepoints

New_feats

(optional) Directly give a feature ndarray. It ignores new_img if given.

Kwargs

Additional keyword values passed to self.predict_proba_from_ndarray

returns

New_lbls

list of labels assigned to new timepoint

New_feats

list of ndarrays containing the updated features

New_max

updated max cell label assigned

predict_proba_from_ndarray(array_3d, model=None, boolean=False, max_distance=None)

input

Array_3d

(ncells_tp1, ncells_tp2, out_feats) ndarray

Model

str, {‘model’, ‘bak_model’} can force a unique model instead of an ensemble

Boolean

bool, if False returns probability, if True returns prediction

Max_distance

float Maximum distance (in um) to be considered. If None it uses the instance’s value,

if zero it skips checking distances.

requires :self.model: :self.mainof_ids: list of indices corresponding to the main model’s features :self.bakof_ids: list of indices corresponding to the backup model’s features

returns

(ncells_tp1, ncells_tp2) ndarray with probabilities or prediction

of cell identities depending on “boolean” arg.

probabilities_from_impair(image_t1, image_t2, kwargs_feat_calc={}, **kwargs)

Convenience function to test tracking between two time-points

Image_t1

np.ndarray containing mask of first time-point

Image_t2

np.ndarray containing mask of second time-point

Kwargs_feat_calc

are passed to self.calc_feats_from_mask calls

Kwargs

are passed to self.predict_proba_from_ndarray

run_tp(masks, state=None, **kwargs)

Assign labels to new masks using a state dictionary.

Parameters

masksnp.ndarray

Cell masks to label.

statet.Dict[str, t.Union[int, list]]

Dictionary containing maximum cell label, and previous cell labels and features for those cells.

kwargskeyword arguments

Keyword arguments passed to self.get_new_lbls

Returns

t.Dict[str, t.Union[t.List[int], t.Dict[str, t.Union[int, list]]]]

New labels and new state dictionary.

Examples

FIXME: Add example beyond the trivial one.

import numpy as np from baby.tracker.core import celltracker from tqdm import tqdm

# overlapping outlines are of shape (t,z,x,y) masks = np.zeros((5, 3, 20, 20), dtype=bool) masks[0, 0, 2:6, 2:6] = true masks[1:, 0, 13:14, 13:14] = true masks[:, 1, 8:12, 8:12] = true masks[:, 2, 14:18, 14:18] = true

# 13um pixel size ct = celltracker(pixel_size=0.185) labels = []

state = none for masks_tp in tqdm(masks):

new_labels, state = ct.run_tp(masks_tp, state=state) labels.append(new_labels)

# should result in state[‘cell_lbls’] # [[1, 2, 3], [4, 2, 3], [4, 2, 3], [4, 2, 3], [4, 2, 3]]

Return type

Dict[str, Union[List[int], Dict[str, Union[int, list]]]]

scale_feats(feats, pixel_size)

input

Feats

np.ndarray (ncells * nfeatures)

Pixel_size

float Value used to normalise the images.

returns Rescaled list of feature values