Research map / Label-efficient and robust learning / Recognition and long-tail
UniCell: Universal Cell Nucleus Classification via Prompt Learning
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 UniCell, an end-to-end DETR-style cell nucleus detection and classification model trained jointly on four pathology datasets with a Dynamic Prompt Module that uses learnable dataset and category prompts to modulate features.
Problem
Pathology datasets for cell nucleus detection and classification use inconsistent label sets, so existing methods train separately per dataset and cannot share knowledge across sources.
Approach
A Dynamic Prompt Module builds dataset prompts (CoOp-style learnable tokens plus dataset name) and a category memory bank, uses an attention mask so each dataset attends only to its own categories, and uses the resulting prompts to enhance multi-scale image features, combined with per-dataset prediction heads in a DETR-based detector with contrastive denoising.
Key results
- UniCell reaches detection F-scores of 0.762 (CoNSeP), 0.847 (MoNuSAC), 0.813 (Lizard), 0.751 (OCELOT), beating the best baseline by 1.3-4.7 points.
- Classification F-scores improve over the best baseline by 8.4, 4.0, 12.2 and 5.6 points on the four benchmarks.
- DPM improves average detection F-score by 1.8 points and average classification F-score by 2.8 points over the variant without DPM (0.793/0.656 vs 0.775/0.628).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Dynamic Prompt Module (ablation) | UniCell · Average across four benchmarks | F_d avg: 0.775 -> 0.793; 1.8% | Helped |
| add the Dynamic Prompt Module (ablation) | UniCell · Average across four benchmarks | F_c̄ avg: 0.628 -> 0.656; 2.8% | Helped |
| add the category memory bank (ablation) | UniCell · Average across four benchmarks | F_c̄ avg: 0.643 -> 0.656 | Helped |
| add the dataset prompts (ablation) | UniCell · Average across four benchmarks | F_c̄ avg: 0.639 -> 0.656 | Helped |
| enhance multi-scale features rather than content queries (ablation) | UniCell DPM · Average across four benchmarks | F_c̄ avg: 0.643 -> 0.656; 1.3% | Helped |
| set the number of Localized Attention Layers to 3 (ablation) | UniCell DPM · Four nucleus benchmarks | classification F-score | Mixed |
| train one model jointly on four datasets using UniCell's prompts rather than separate dataset models | nucleus detection and classification models · CoNSeP, MoNuSAC, Lizard, OCELOT | F-score: 4.7%, 1.9%, 1.4% and 1.3% | Helped |
Surprising findings
- Training baselines jointly on all four datasets (merged labels) often degrades them compared with per-dataset training, e.g. UperNet ConvNeXt on Lizard drops from 0.764 to 0.458 detection F-score.
- Pretraining on binary detection-classification or merging categories into one head did not improve over individual training in the authors' ablation.
Limitations
- Relies on centroid-only annotations; boundary segmentation is not addressed.
- Class imbalance worsens after merging datasets with non-overlapping label sets.
Experimental setup
Datasets: CoNSeP MoNuSAC Lizard OCELOT
Benchmarks: CoNSeP MoNuSAC Lizard OCELOT
Metrics: detection F-score classification F-score average classification F-score
Baselines: HoverNet Cellpose Omnipose UperNet ConvNeXt
Builds on: DETR DINO CoOp Swin Transformer
Closest papers by meaning
- Incremental Nuclei Segmentation from Histopathological Images via Future-class Awareness and Compatibility-inspired Distillation (CVPR 2024 · similarity 0.85)
- Cell Graph Transformer for Nuclei Classification (AAAI 2024 · similarity 0.85)
- MAPLE: Multi-scale Attribute-enhanced Prompt Learning for Few-shot Whole Slide Image Classification (NeurIPS 2025 · similarity 0.84)
- Are nuclear masks all you need for improved out-of-domain generalisation? A closer look at cancer classification in histopathology (NeurIPS 2024 · similarity 0.84)
- Memory-Efficient Prompt Tuning for Incremental Histopathology Classification (AAAI 2024 · similarity 0.83)
- xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.