# An objective comparison of cell-tracking algorithms

##### Posted on Aug 21, 2020
Tags: Single Cell

## Introduction

A combined report on the results of three editions of the Cell Tracking Challenge.

• 21 participating algorithms
• 13 data sets from various microscopy modalities

Cell migration and proliferation are two important processes in normal tissue development and disease, and optical microscopy remains the most appropriate imaging modality (??) for visualizing these processes.

Image techniques:

• phase contrast (PhC) or differential interference contrast (DIC) microscopy: make cells visible without the need of exogeneous markers
• fluorescence microscopy: relies on fluorescent reporters to specifically label cell components such as nuclei (细胞核), cytoplasm (细胞质), or membranes (细胞膜). These labeled structures are then imaged in two or three dimensions by various imaging modalities, including widefield, confocal (共焦的), multiphoton (多光子) or light-sheet (??) fluorescent microscopy.

Tasks: accurately delineating (that is, segmenting) cell boundaries and tracking cell movements over time, providing information about their velocities and trajectories, and detecting cell-lineage changes as a result of cell division or cell death.

For cell segmentation, creating a ‘taxonomy of methods’ is not a straightforward process, as state-of-the-art methods usually combine different strategies to achieve improved results. Classify existing algorithms by three criteria:

• the principle on which cells are detected, e.g., by finding uniform areas, boundaries or at very low resolution by simply finding bright spots and maxima.
• the image features that are computed to achieve the cell segmentation. These can be simple pixel or voxel intensities, their local averages, or more complex local image descriptors of shapes or textures.
• distinguish the segementation method itself that implements the principle using the features. The methods ranges from simple methods like thresholding, hysteresis (滞后) thresholding, edge detection and shape matching to more sophisticated approaches like region growing, machine learning, and energy minimization.

Cell-tracking methods can be broadly categorized into two groups:

• tracking by contour evolution methods: start by segmenting the cells in the first frame of a video and then evolve their contours in consecutive frames, thereby solving the segmentation and tracking tasks simultaneously, one step at a time, under the essential assumption of unambiguous, spatiotemporal overlap between the corresponding cell regions.
• tracking by detection methods: start by segmenting the cells in all frames of a video and later, using mostly probabilistic frameworks, establish temporal associations between the segmented cells. This can be done by either using a two-frame or multiframe sliding window, or even for all frames at once.

## Data sets and ground truth

13 data sets:

• 11 consist of contrast enhancing (PhC, DIC) or fluorescence (wide-field, confocal, light sheet) microscopy recordings of live cells and organisms in 2D or 3D.
• the other two data sets are synthetic, generated using a cell simulator that produces realistic 2D and 3D renderings of chromatin-stained live cells.

## Quantitative performance criteria

1. the segmentation and tracking accuracy from the computer science point of view
• segmentation accuracy measure (SEG)
• tracking accuracy measure (TRA)
• overall performance (OP): $\text{OP} = (\text{SEG}{avg} + \text{TRA}{avg}) / 2$
2. the biological relevance of the obtained tracking results
• complete tracks (CT): measure the fraction of ground truth cell tracks that a given method is able to reconstruct in their entirety, from the frame they appear in to the frame they disappear from. It is especially relevant when a perfect reconstruction of the cell lineages is required.
• track fraction (TF): for all detected tracks, averages the fraction of the longest continuously matching algorithm-generated tracklet with respect to the reference track. It can be interpreted as the fraction of an average cell’s trajectory that an algorithm reconstruct correctly once the cell has been detected.
• branching correctness (BC): measures how efficient a method is at detecting division events
• cell cycle accuracy (CCA): measures how accurate an algorithm is at correctly reconstructing the length of cell cycles (that is, the time between two consecutive divisions)
3. the practical usability of the methods
• the number of tunable parameters (NP) a user is required to manually set, excluding parameters visible only to developers. Generally, a lower number of tunable parameters indicates a more usable algorithm.
• generalizability (GP): quantify how stable an algorithm is when being applied with the same parameter configuration to new videos acquired under otherwise unchanged imaging conditions.
• execution time (TIM)

Published in categories Note