[Paper Review] Tracking Tetrahymena Pyriformis Cells using Decision Trees

Quan Wang, Yan Ou+3 more|arXiv (Cornell University)|Jul 13, 2012

Microbial Community Ecology and Physiology4 references3 citations

TL;DR

This paper proposes a real-time cell tracking method for fast-moving Tetrahymena pyriformis cells using decision trees to classify cell associations across frames. By computing a 23-dimensional feature difference vector from spatial, intensity, and composite features, and applying a modified Hungarian algorithm to resolve assignments, the method achieves a low 1.37% misclassification rate at tree depth 8, enabling robust tracking under occlusion and appearance variation.

ABSTRACT

Matching cells over time has long been the most difficult step in cell tracking. In this paper, we approach this problem by recasting it as a classification problem. We construct a feature set for each cell, and compute a feature difference vector between a cell in the current frame and a cell in a previous frame. Then we determine whether the two cells represent the same cell over time by training decision trees as our binary classifiers. With the output of decision trees, we are able to formulate an assignment problem for our cell association task and solve it using a modified version of the Hungarian algorithm.

Motivation & Objective

• Address the challenge of accurate cell tracking in live-cell microscopy, particularly for fast-moving and morphologically similar cells.
• Overcome difficulties in cell association due to occlusion, overlapping cells, and appearance changes over time.
• Develop a real-time, scalable solution for multi-cell control systems requiring precise positional feedback.
• Formulate cell association as a binary classification problem using learned decision trees on feature difference vectors.
• Integrate the classifier output into a modified Hungarian algorithm to solve the global assignment problem efficiently.

Proposed method

• Perform background subtraction using median filtering over the first 20 frames to generate a stable background image.
• Extract 23-dimensional feature difference vectors between candidate cells in consecutive frames, combining spatial, intensity, and composite features.
• Train decision trees (T1 and T2) on labeled cell pairs to classify whether two cells in adjacent frames represent the same cell (f(v) = 1) or not (f(v) = 0).
• Use a modified Hungarian algorithm that handles unmatched cells and overlapping regions by introducing dummy columns for 'new' and 'overlapping' associations.
• Update cell lists dynamically: assign active, new, occluded, or out-of-bound statuses based on association results and spatial proximity.
• Apply constant-velocity prediction to estimate the expected position of a cell in the next frame, used in the first feature of the difference vector.

Experimental results

Research questions

• RQ1 Can decision trees effectively classify cell associations in time-lapse microscopy despite appearance variation and occlusion?
• RQ2 How well do feature difference vectors composed of spatial, intensity, and composite features perform in distinguishing true cell matches from false positives?
• RQ3 To what extent does the modified Hungarian algorithm improve global assignment accuracy when combined with decision tree classification?
• RQ4 How sensitive is the tracking performance to hyperparameters such as tree depth, feature discretization, and distance thresholds?
• RQ5 Can this method achieve real-time performance and robustness in tracking fast-moving, morphologically similar cells under complex interactions?

Key findings

• The decision tree of depth 8 achieved the lowest misclassification rate of 1.37% on the test set, outperforming shallower and deeper trees.
• All 23 entries in the feature difference vector were found to be statistically significant and contributed to classification performance.
• The modified Hungarian algorithm successfully resolved global assignments, including handling new cells and overlapping regions, improving tracking consistency.
• The method demonstrated robustness in tracking scenarios with high cell density, occlusions, and rapid movements, as shown in Figure 3.
• Tracking results were stable under minor parameter variations, indicating low sensitivity to hyperparameter tuning.
• The system achieved real-time performance suitable for integration into feedback-controlled systems for magnetotactic T. pyriformis cells.