TL;DR
This paper introduces a novel multi-worm tracking method for C. elegans that combines an improved skeletonization algorithm for handling collisions with a network flow model for trajectory reconstruction. This approach effectively segments and tracks individual worms even when they overlap, achieving high accuracy in complex scenarios.
What Problem Does This Paper Address?
Tracking multiple C. elegans simultaneously is crucial for understanding their behavior in neuroscience and genetics research. However, a significant challenge arises when worms collide and overlap, making it difficult to distinguish them and maintain their individual trajectories. Existing methods often struggle with these collisions, leading to fragmented or incorrect tracks, and some deep learning approaches require extensive labeled data. This paper aims to provide a robust and accurate solution for multi-worm tracking, specifically addressing the problem of worm collisions.
Methods
The proposed method is divided into two main stages: intra-trajectory and inter-trajectory.
- Intra-trajectory Stage:
- Worm Detection: Standard morphological methods are used to detect worm responses in each frame.
- Collision Detection: An area threshold (
Tarea) is used to distinguish between collided and non-collided worms. - Non-Collided Worms: Trajectory fragments are generated by matching the center skeleton points of worms in adjacent frames based on spatio-temporal features (distance and angle).
- Collided Worms: An improved skeleton method is introduced. This method first removes branchpoints from the traditional skeleton to obtain individual worm skeletons. Then, a dynamic matching method, informed by skeleton length, is used to effectively segment and separate the skeletons of colliding worms. This is a key novelty, as traditional skeletonization struggles with overlaps.
- Inter-trajectory Stage:
- Network Flow Model: The trajectory fragments generated from both collided and non-collided worms are treated as vertices in a network flow model.
- Minimum Cost Flow: A minimum-cost flow algorithm is employed to solve this model. The cost function incorporates features like position, appearance (using Bhattacharyya distance for similarity), and velocity (using cosine similarity) to find the optimal multi-worm trajectories. This approach leverages the strengths of both individual worm features and their relationships across frames.
Datasets
The method was evaluated using video data of C. elegans populations at three distinct ages: L4, young adult, and D1. The paper does not mention the creation of new benchmark datasets but rather uses existing video data for evaluation.
Key Insights and Ideas
- Skeletonization for Collision Resolution: The core insight is that by intelligently processing skeleton information, specifically by removing branchpoints and using dynamic matching, individual worm shapes can be disentangled even during collisions. This is a significant improvement over methods that fail when worms overlap.
- Network Flow for Global Optimization: Treating trajectory fragments as vertices in a network flow model allows for a global optimization approach to reconstruct complete trajectories. This is more robust than purely local frame-to-frame matching, especially in complex scenarios.
- Feature Integration for Cost Function: Combining appearance and motion features in the cost function of the network flow model provides a richer representation for data association, leading to more accurate tracking.
Key Results
The proposed method achieved promising results:
- MOTA (Multiple Object Tracking Accuracy): Between 0.86 and 0.92 across different C. elegans ages.
- MOTP (Multiple Object Tracking Precision): Between 0.78 and 0.83.
- Comparative Performance: In a comparison with the "deep-worm tracker" [10], the proposed method showed competitive or superior MOTA scores on several videos, particularly in scenarios with a higher number of worms or complex interactions. For instance, on "Worm3" (9-10 worms), the proposed method achieved 91.20% MOTA compared to 85.66% for the deep-worm tracker.
Limitations and Future Work
- Limitations:
- The primary source of error is attributed to the worm detection algorithm, which uses adaptive thresholding and can miss smaller worm blobs or be affected by background impurities.
- Errors can also arise from the network flow method when approaching worms have similar appearance and speed, leading to potential judgment errors.
- The current method uses position, appearance, and motion speed as features.
- Future Work:
- Advanced Worm Detection: Employing more sophisticated detection methods, such as combining optimization theory with probabilistic detectors or using learning-based detectors (though this might require more computational resources).
- Enhanced Feature Set: Incorporating additional worm features into the network flow cost function, such as amplitude and frequency of movement, to improve tracking accuracy when appearance and speed are insufficient for differentiation.
- The authors aim to optimize these aspects to further enhance tracking performance.
Summary updated