Higher-Order Cell Tracking Transformer

Cell tracking in live-imaging microscopy is fundamentally difficult because cells often appear visually identical and undergo complex division events. Traditional approaches typically construct a candidate graph of cell detections and use graph neural networks (GNNs) to link them. However, these methods often struggle because cell divisions create "connected manifolds" in node embedding spaces, and the underlying candidate graphs are non-homophilic—meaning that the graph's topology provides almost no useful information for distinguishing correct links from incorrect ones.

The authors introduce the Higher-Order Cell Tracking Transformer (HOCT), an edge-centric architecture designed to resolve these structural obstacles. Instead of focusing on nodes, HOCT treats candidate cell links as independent tokens. It uses a two-stage Transformer: first, a node-based stage contextualizes cell detections, and second, an edge-based stage refines candidate links through self-attention. Crucially, HOCT incorporates a geometric prior based on the 3D line-to-line distance between candidate edges. This allows the model to attend to spatial relationships between potential trajectories rather than relying on the uninformative adjacency of the candidate graph. The final tracking solution is recovered using a two-pass integer linear programming (ILP) solver that enforces biological constraints like cell division and appearance.

HOCT achieves state-of-the-art performance on the Cell Tracking Challenge and a bacteria division benchmark without requiring deep pre-trained image encoders. The authors demonstrate that the edge-centric design significantly outperforms topology-driven GNNs, which are hindered by the non-homophilic nature of tracking graphs. Furthermore, HOCT shows superior efficiency in human-in-the-loop settings; it reduces tracking errors by 59% with only 400 annotations, outperforming LoRA fine-tuning of standard node-based transformer baselines. The inclusion of variable appearance probabilities and a two-pass tracklet solver further improves the robustness of the lineage reconstruction.