Research map / Image generation / Fast diffusion sampling
From Representation Space to Prognostic Insights: Whole Slide Image Generation with Hierarchical Diffusion Model for Survival Prediction
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 WSI-Diffusion, a two-stage hierarchical DDPM that generates missing WSI representations (WSI-level then patch-level, with tissue-type distribution-aware sampling) to augment patients with few WSIs for survival prediction.
Problem
Survival prediction from whole slide images is biased against patients with few WSIs because WSI preparation yields variable WSI counts per patient, and existing augmentation methods only modify existing patches rather than creating new WSIs.
Approach
Generate entirely new WSI representations in feature space via a hierarchical diffusion process: a WSI-level diffuser produces a WSI feature (patch-consistency constraint) and tissue-type distribution, which conditions a patch-level diffuser that samples class-conditioned patch embeddings; generated WSIs are added as supplementary data for a light C-MIL-based Cox survival model.
Key results
- WSI-Diffusion reaches 0.751 C-index on NLST, 0.714 on LUSC, 0.668 on LUAD, 0.660 on BRCA and 0.601 on BLCA, outperforming all listed baselines.
- Averages a 3.2% C-index and 2.6% STAGE-5 gain over competing methods across five datasets (5-fold CV).
- On NLST, the performance gap between low- and high-WSI-count patients shrinks from 0.042 without generation to 0.014 with generated WSIs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add WSI-Diffusion-generated WSIs | HIPT survival prediction · NLST, LUSC, LUAD, BRCA, BLCA | C-index | Helped |
| add WSI-Diffusion-generated WSIs | HIPT survival prediction · NLST, LUSC, LUAD, BRCA, BLCA | STAGE-5 | Helped |
| add generated WSIs | survival prediction for patients grouped by WSI count · NLST, LUAD, LUSC | C-index gap between low- and high-WSI groups: 0.042 -> 0.014; 0.027 -> 0.016; 0.024 -> 0.013 | Helped |
| add generated WSIs | survival prediction for high-WSI-count patients · NLST, LUAD, LUSC | C-index | Helped |
| add WSI generation (ablation) | WSI-Diffusion · NLST and LUSC | C-index: 0.698.030 -> 0.751.008; 0.668.019 -> 0.714.012 | Helped |
| add the WSI-level diffuser (ablation) | WSI-Diffusion · NLST and LUSC | C-index: 0.685.029 -> 0.751.008; 0.663.021 -> 0.714.012 | Helped |
| add tissue type distribution constraint (ablation) | WSI-Diffusion · NLST and LUSC | C-index: 0.709.022 -> 0.751.008; 0.690.023 -> 0.714.012 | Helped |
| add patch consistency constraint (ablation) | WSI-Diffusion · NLST and LUSC | C-index: 0.725.013 -> 0.751.008; 0.700.023 -> 0.714.012 | Helped |
| increase diffusion steps from 10 to 30 (ablation) | WSI-Diffusion · NLST, LUSC, LUAD | C-index: 0.726.012 -> 0.751.008; 0.682.030 -> 0.714.012; 0.632.015 -> 0.668.012 | Helped |
| increase diffusion steps from 30 to 50 (ablation) | WSI-Diffusion · NLST, LUSC, LUAD | C-index: 0.751.008 -> 0.750.009; 0.714.012 -> 0.712.020; 0.668.012 -> 0.665.011 | No clear effect |
| generate WSIs for all patients rather than only patients with fewer than three WSIs (ablation) | WSI-Diffusion · NLST, LUSC, LUAD | C-index: 0.751.008 vs 0.732.015; 0.714.012 vs 0.682.030; 0.668.012 vs 0.632.020 | Helped |
| generate the same total number of WSIs for each patient (ablation) | WSI-Diffusion · NLST, LUSC, LUAD | C-index: 0.751.008 vs 0.689.042; 0.714.012 vs 0.657.033; 0.668.012 vs 0.601.030 | Hurt |
Surprising findings
- Generating only with the patch-level diffuser (no WSI-level constraints) does not outperform the no-generation baseline, suggesting unconstrained generated patches add noise.
Limitations
- Tissue-type distribution relies on HoVer-Net nucleus-type majority voting, limiting the granularity of tissue categories.
- Performance plateaus with more generated WSIs and longer diffusion time steps, and partial generation for only low-count patients is less effective.
- Subtype labels are absent in the datasets, so the authors plan pseudo-labeling via clustering, which may reduce explainability.
Experimental setup
Datasets: NLST TCGA-LUSC TCGA-LUAD TCGA-BRCA TCGA-BLCA PanNuke
Benchmarks: NLST TCGA-LUSC TCGA-LUAD TCGA-BRCA TCGA-BLCA
Metrics: C-index STAGE-5
Baselines: WSISA DeepAttnMISL HIPT SeTranSurv ESAT LongViT Prov-GigaPath AugDiff
Builds on: DDPM AugDiff C-MIL HIPT HoVer-Net
Closest papers by meaning
- Multimodal Prototyping for cancer survival prediction (ICML 2024 · similarity 0.87)
- Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology (ICLR 2025 · similarity 0.86)
- Leveraging Tumor Heterogeneity: Heterogeneous Graph Representation Learning for Cancer Survival Prediction in Whole Slide Images (NeurIPS 2024 · similarity 0.86)
- Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction (CVPR 2024 · similarity 0.85)
- Toward a Well-Calibrated Discrimination via Survival Outcome-Aware Contrastive Learning (NeurIPS 2024 · similarity 0.84)
- Tumor Micro-environment Interactions Guided Graph Learning for Survival Analysis of Human Cancers from Whole-slide Pathological Images (CVPR 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.