CytoTracker: Multi-Cell Tracking with Diffusion-Guided Cell Association

Komal Kumar, Noa Novershtern, Jacob hanna, Hisham Cholakkal

Keywords: cells detection, lineage tracking, cell association, diffusion motion.

Abstract:

Cell tracking is a challenging task in microscopic image analysis that enables the study of cellular behaviors and interactions over time. It requires tracking hundreds of nearly indistinguishable cells that move dynamically and may undergo growth or division, complicating the tracking process. As a result, state-of-the-art cell tracking methods often rely on ground truth segmentation mask-derived cell-level features, even during inference, which are typically unavailable in real-world applications.
In this work, we introduce CytoTracker, an end-to-end framework for cell tracking that integrates three key modules: a cell detector, a diffusion model-based motion prediction network, and a transformer-based association network. Additionally, we developed the CytoEmbedding attention fusion block to effectively extract cell-level features from images, improving tracking. Our method accurately detects cells in each frame, predicts nonlinear motion, and robustly associates them across time, even in the presence of cell division. Experimental results on microscopy datasets demonstrate that CytoTracker achieves performance comparable to state-of-the-art approaches that require ground truth masks during inference, without such costly segmentation mask annotations.

Primary Subject Area: Foundation Models

Secondary Subject Area: Application: Other

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 253