Research map / Graph learning / Graph self-supervised learning

Higher-Order Cell Tracking Transformer

NeurIPS 2026 · Graph learning › Graph self-supervised learning · approach: Brain and biological graph learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HOCT, an edge-centric Transformer for cell tracking that lets candidate links attend to each other with a learnable line-to-line distance bias, a multi-frame parental softmax and an extended ILP, achieving SOTA on the Cell Tracking Challenge and a bacteria benchmark without pretrained image encoders.

Problem

Linking cell detections across time (including divisions) for lineage reconstruction; node-embedding and GNN approaches on candidate tracking graphs suffer from division-induced entanglement and non-homophilic topology.

Approach

Operate in an edge embedding space: candidate links are tokens that attend to one another with RoPE plus per-head learnable attractive/repulsive line-to-line distance biases, instead of node-topology aggregation; combined with a multi-gap parental softmax and two-pass tracklet ILP.

Key results

• Official CTC hidden-test leaderboard: HOCT ranks first in overall CLB 0.930, LNK 0.985, BIO 0.875 (Generalizability).
• Bacteria benchmark: HOCT achieves AOGM 6.36 ± 1.35 with zero deletion errors using 19-dim hand-crafted features, vs 6.53 for best Trackastra variant (SAM2.1 features).
• Incremental correction: HOCT uncertainty sampling reduces AOGM by 58.6% after 400 annotations vs 6.75% for LoRA + unsup Trackastra+Attrackt.

Surprising findings

• Candidate tracking graphs have adjusted homophily near zero (H_adj ≈ 0.01), so topology carries no label signal for GNN aggregation.
• A logistic regression head on frozen HOCT edge features outperforms LoRA fine-tuning of a node-based Transformer in human-in-the-loop correction.

Limitations

• Edge-centric attention increases computational and memory cost (O(|E|^2)).
• Very large fields of view with thousands of cells per frame require tiling.
• Dataset-specialist methods outperform HOCT on a few datasets (e.g., U373, Fluo2D SIM+).

Experimental setup

Datasets: Cell Tracking Challenge (16 datasets) bacteria tracking benchmark

Benchmarks: Cell Tracking Challenge hidden test set bacteria tracking benchmark bacteria validation incremental correction

Metrics: CLB LNK BIO AOGM AOGM-A AOGM-D Division F1

Baselines: Trackastra Edge-Transformer GAT FAGCN Attrackt Baxter Algorithm MAMHT TrackTour ByoTrack zTrack MAGIK Medical SAM2

Builds on: Trackastra RoPE parental softmax Attrackt focal loss

Closest papers by meaning

• Cell Graph Transformer for Nuclei Classification (AAAI 2024 · similarity 0.83)
• Bridging Protein Sequences and Microscopy Images with Unified Diffusion Models (ICML 2025 · similarity 0.82)
• Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning (ICML 2025 · similarity 0.82)
• Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction (NeurIPS 2026 · similarity 0.82)
• ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.82)
• TopoCellGen: Generating Histopathology Cell Topology with a Diffusion Model (CVPR 2025 · similarity 0.81)