Enhancing Modeling Capability of Graph Neural Network for Cell Tracking via Topological Information Exploitation

Cell tracking entails constructing trajectories from microscopic videos by associating instances of the same cell across sequential frames. Precise trajectory construction is essential for understanding disease progression and evaluating therapeutic interventions. To minimize reliance on expert input, various automated cell tracking methodologies employing deep learning models have been developed, predominantly emphasizing visual similarity and temporal dependencies. Prior research has highlighted that the rich topological characteristics of cells provide significant insights into multi-cellular developmental dynamics. Nonetheless, current deep learning models for cell tracking do not fully exploit this topological information. To address this limitation, we introduce a Topological Graph Neural Network for Cell Tracking (TGNN-CELL), which harnesses local topological features to enhance trajectory prediction accuracy. TGNN-CELL generates latent representations of the Persistence Diagram (PD) for each node’s neighborhood, capturing multi-scale local topology. This local topological representation is incorporated into cellular embeddings to enhance cell connection prediction capability of GNN. Empirical evaluations demonstrate that integrating topological information enhances model performance in cell tracking tasks, with TGNN-CELL surpassing existing baselines by 4% to 7% in target effectiveness across three conditions on the benchmark dataset C2C12. Our approach effectively utilizes topological data to augment the modeling capabilities of the GNN framework for cell tracking. This study establishes a foundation for future investigations into leveraging cellular topology to improve the precision of deep learning models in various biological applications.