traccuracy.metrics._divisions

This submodule classifies division errors in tracking graphs

Each division is classified as one of the following: - true positive - false positive - false negative

These functions require two TrackingGraph objects and a mapper between nodes in the two graphs. Divisions are identified as correct if both the parent and daughter nodes match between the GT and predicted graph.

Temporal tolerance for correct divisions is implemented to allow for cases in which the exact frame that a division event occurs is somewhat arbitrary due to a high frame rate or variable segmentation or detection. Consider the following graphs as an example:

G1
                            2_4
1_0 -- 1_1 -- 1_2 -- 1_3 -<
                            3_4
G2
              2_2 -- 2_3 -- 2_4
1_0 -- 1_1 -<
              3_2 -- 3_3 -- 3_4

After classifying basic division errors, we consider all false positive and false negative divisions. If a pair of errors occurs within the specified frame buffer, the pair is considered a true positive division if the parent nodes and daughter nodes match. We determine the “parent node” of the late division, e.g. node 1_3 in graph G1, by traversing back along the graph until we find the node in the same frame as the parent node of the early division. We repeat the process for finding daughters of the early division, by advancing along the graph to find nodes in the same frame as the late division daughters.

Module Contents

Classes

DivisionMetrics

Computes division summary metrics with an optional frame tolerance.

Attributes

logger

traccuracy.metrics._divisions.logger
class traccuracy.metrics._divisions.DivisionMetrics(max_frame_buffer: int = 0, zero_division: float = np.nan)[source]

Computes division summary metrics with an optional frame tolerance.

Computes the following metrics: - Division Recall - Division Precision - Division F1 Score (also Branching Correctness) - Mitotic Branching Correctness: TP / (TP + FP + FN) as defined by Ulicna, K., Vallardi, G., Charras, G. & Lowe, A. R. Automated deep lineage tree analysis using a Bayesian single cell tracking approach. Frontiers in Computer Science 3, 734559 (2021).

These metrics are written assuming that the ground truth annotations are dense. If that is not the case, interpret the numbers carefully. Consider eliminating metrics that use the number of false positives.

Parameters:

  • max_frame_buffer (int, optional) – Maximum value of frame buffer to use in correcting shifted divisions. Divisions will be evaluated for all integer values of frame buffer between 0 and max_frame_buffer

  • zero_division (float, optional) – Value to return for metrics that result in a 0/0 division. Defaults to np.nan. Set to 0.0 to return 0 and raise a warning instead, similar to scikit-learn’s zero_division parameter.

property info: dict[str, Any]

Dictionary with Metric name and any parameters

compute(matched: traccuracy.matchers._matched.Matched, override_matcher: bool = False, relax_skips_gt: bool = False, relax_skips_pred: bool = False) → traccuracy.metrics._results.Results

The compute methods of Metric objects return a Results object populated with results and associated metadata

Parameters:

  • matched (traccuracy.matchers.Matched) – Matched data object to compute metrics on

  • override_matcher (bool) – If True, the metric will not validate the matcher type

  • relax_skips_gt (bool) – If True, the metric will check if skips in the ground truth graph have an equivalent multi-edge path in predicted graph

  • relax_skips_pred (bool) – If True, the metric will check if skips in the predicted graph have an equivalent multi-edge path in ground truth graph

Returns:

Object containing metric results

and associated pipeline metadata

Return type:

traccuracy.metrics._results.Results