Research map / Image generation / Fast diffusion sampling
TopoCellGen: Generating Histopathology Cell Topology with a Diffusion Model
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 TopoCellGen, a DDPM-based cell-layout generator that adds cell-count, intra-class and inter-class persistent-homology losses, and a TopoFD metric for evaluating topological fidelity of generated multi-class cell layouts.
Problem
Generating realistic multi-class cell spatial layouts for histopathology that preserve intra-class distributions, inter-class topological relationships, and accurate cell counts, which existing image-level diffusion models and FID-based evaluation do not capture.
Approach
Integrates persistent-homology-based topological losses (intra-class and aggregated inter-class, computed on distance-transform maps) and a straight-through-estimator cell counting loss into diffusion training for cell layouts, and introduces TopoFD, a Fréchet distance over persistence-landscape statistics.
Key results
- On BRCA-M2C, TopoCellGen reaches FID 0.005 and TopoFD 69.354 versus 0.263 and 97.584 for Spatial Diffusion, and TCE 5.192 versus 20.806.
- On Lizard, TopoCellGen reduces TCE to 11.590 and TopoFD to 31.607 compared with the best baseline TMCCG (22.604 TCE, 63.120 TopoFD).
- Adding synthetic image-layout pairs as augmentation improves mean detection-classification F1 with UNet from 0.604 to 0.678 on BRCA-M2C.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the cell counting loss Lcount (ablation) | TopoCellGen · BRCA-M2C | TCE: 22.465 -> 12.253 | Helped |
| Add the intra-class spatial consistency loss Lintra (ablation) | TopoCellGen · BRCA-M2C | FID: 1.150 -> 0.621 | Helped |
| Add the inter-class structural loss Linter (ablation) | TopoCellGen · BRCA-M2C | TopoFD: 133.012 -> 126.742 | Helped |
| Use all three losses together (ablation) | TopoCellGen · BRCA-M2C | TopoFD: 133.012 -> 69.354 | Helped |
| Set λc, λintra, and λinter to 5e-4 rather than 0.005 (ablation) | TopoCellGen · BRCA-M2C | FID: 0.289 -> 0.005 | Helped |
| Augment UNet training with TopoCellGen synthetic image-layout pairs instead of Spatial Diffusion pairs | UNet · BRCA-M2C cell detection and classification | Mean F1-score: 0.666 ± 0.007 -> 0.678 ± 0.004 | Helped |
| Augment MCSpatNet training with TopoCellGen synthetic image-layout pairs instead of Spatial Diffusion pairs | MCSpatNet · BRCA-M2C cell detection and classification | Mean F1-score: 0.669 ± 0.006 -> 0.684 ± 0.004 | Helped |
Surprising findings
- The cell counting loss alone reduces total count error but the intra-class loss alone yields the lowest TopoFD among single losses in some settings, while inter-class loss has little effect on counts.
Limitations
- Evaluation limited to two nuclei datasets and three to four cell types per dataset.
- FID feature extractors are custom-trained per dataset, limiting comparability.
- Biological plausibility assessed by a single pathologist.
Experimental setup
Datasets: BRCA-M2C Lizard
Benchmarks: BRCA-M2C Lizard
Metrics: FID total count error TopoFD MMD F1-score
Baselines: ADM TMCCG Spatial Diffusion MCSpatNet
Builds on: DDPM TopoGAN TopoDiffusionNet guided diffusion
Closest papers by meaning
- TopoDiffusionNet: A Topology-aware Diffusion Model (ICLR 2025 · similarity 0.86)
- Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.85)
- STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation (NeurIPS 2026 · similarity 0.84)
- Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction (NeurIPS 2026 · similarity 0.84)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.84)
- MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation (NeurIPS 2025 · 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.