Versatile multiple object tracking in sparse 2D/3D videos via deformable image registration.

Area of Science:

• Computational Biology
• Bioimage Analysis
• Machine Learning

Background:

• Automated and semi-automated multiple object tracking (MOT) tools are crucial for analyzing large biological datasets.
• Existing MOT methods often lack generalizability across different datasets or require extensive training data.
• Tracking fluorescent sources in deforming tissues presents challenges due to sparse data and lack of unique features.

Purpose of the Study:

• To develop a versatile and generalizable image registration framework for semi-supervised MOT in biological videos.
• To address the limitations of existing MOT methods in handling complex biological imaging data.
• To provide an accessible tool for researchers tracking objects in 2D and 3D biological videos.

Main Methods:

• Proposed ZephIR, an image registration framework leveraging spatial transformer networks for semi-supervised MOT.
• Incorporated adjustable parameters for spatial (sparsity, texture, rigidity) and temporal priors to enhance generalizability.
• Developed an open-source package with a web-based graphical user interface for interactive user input.

Main Results:

• Demonstrated ZephIR's accuracy and versatility across diverse biological applications.
• Successfully tracked body parts of a behaving mouse.
• Accurately tracked neurons in the brain of a freely moving C. elegans.

Conclusions:

• ZephIR offers a robust and adaptable solution for semi-supervised MOT in challenging biological imaging scenarios.
• The framework's ability to generalize across different biological systems and its user-friendly interface facilitate broader adoption.
• ZephIR significantly advances the capabilities for analyzing dynamic biological processes from video data.