Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The Higher-Order Cell Tracking Transformer (HOCT) addresses the challenges of reconstructing cell lineages from live-imaging microscopy by employing an edge-centric architecture that leverages a 3D geometric prior to improve candidate cell link associations. This method effectively resolves issues related to entangled lineage paths and random label agreement in candidate tracking graphs, which have hindered the performance of existing graph neural network approaches. Evaluated against benchmarks, HOCT achieves state-of-the-art performance, reducing tracking errors by 59% with minimal annotations, significantly outperforming traditional fine-tuning methods.
Tracking cell lineages just got a major upgrade—HOCT slashes tracking errors by 59% with just 400 annotations, outpacing existing transformer models.
Reconstructing lineages from live-imaging microscopy requires linking cell detections across time, including through cell divisions. A common approach is to construct a candidate graph and associate cell segmentations (nodes) across frames. However, these and other existing methods overlook two structural obstacles in candidate tracking graphs: (i) cell divisions entangle distinct lineage paths in the node embedding space, and (ii) edges sharing a node have near-random label agreement, so the candidate-graph topology carries no useful information for graph neural networks to aggregate. We propose the \textbf{Higher-Order Cell Tracking Transformer} (HOCT), an edge-centric architecture in which candidate cell links attend to one another under a 3D geometric prior, resolving both issues. Evaluated on the Cell Tracking Challenge and a bacteria division benchmark, HOCT achieves state-of-the-art results without deep pre-trained image encoders. Moreover, the proposed approach is easier to fine-tune, quickly reducing tracking errors by 59% with 400 annotations in a human-in-the-loop setting, outperforming LoRA fine-tuning of competing transformer baselines (6.75% improvement).