Research map / Image generation / Fast diffusion sampling
Dynamic Entity-Masked Graph Diffusion Model for Histopathology Image Representation 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 H-MGDM, a self-supervised histopathology representation learner that pre-trains a latent graph diffusion model on superpixel entity graphs, using complementary visible and masked subgraphs as condition and reconstruction target.
Problem
Mask-based SSL for histopathology patches masks grid tiles and ignores spatial interactions among tissue entities, while natural-image pretrained models suffer domain gap and annotation scarcity limits retraining.
Approach
Convert each patch into a superpixel entity graph encoded in latent space, randomly split it into complementary visible and masked subgraphs, and reconstruct the masked subgraph with a latent graph diffusion decoder conditioned on visible encoder features under dynamic noise levels.
Key results
- On Komura et al. classification, H-MGDM reaches 82.06% ACC and 72.41% F1, above the best baseline DiffMAE (78.14%/70.23%).
- On IBD classification, H-MGDM reaches 86.23% ACC vs 84.58% for DiffMAE.
- On TCGA-KIRC survival with PatchGCN, H-MGDM achieves a C-index of 71.17 vs 68.29 for DiffAE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add edge latents (ablation) | H-MGDM · Komura et al., PANDA, and IBD | ACC (%): 80.91±1.92 -> 82.06±1.17; 72.39±1.29 -> 74.51±1.13; 84.11±1.72 -> 86.23±2.31 | Helped |
| add skip connections (ablation) | H-MGDM · Komura et al., PANDA, and IBD | ACC (%): 79.29±1.27 -> 82.06±1.17; 70.98±1.75 -> 74.51±1.13; 82.35±2.32 -> 86.23±2.31 | Helped |
| use dynamic rather than fixed noise intensity (ablation) | H-MGDM · Komura et al., PANDA, and IBD | ACC (%): 79.70±1.14 -> 82.06±1.17; 72.52±1.65 -> 74.51±1.13; 83.58±1.69 -> 86.23±2.31 | Helped |
| increase mask ratio from 0.5 to 0.6 (ablation) | H-MGDM · Komura et al. | ACC: 78.38 -> 82.06 | Helped |
| increase mask ratio from 0.6 to 0.7 (ablation) | H-MGDM · PANDA | ACC: 72.49 -> 74.51 | Helped |
| increase mask ratio from 0.4 to 0.5 (ablation) | H-MGDM · IBD | ACC: 84.51 -> 86.23 | Helped |
| align vertex and edge attributes separately (NtoN & EtoE) rather than cross-aligning them (NtoE & EtoN) (ablation) | H-MGDM decoder conditioning · PANDA | ACC: 73.82 -> 74.51 | Helped |
| use heterogeneous vertex-edge conditioning (NtoE & EtoN) rather than aligned conditioning (NtoN & EtoE) (ablation) | H-MGDM decoder conditioning · PANDA | RMSE: 0.141 -> 0.127 | Helped |
| use H-MGDM features instead of the strongest listed baseline | classification models · Komura et al., PANDA, and IBD | ACC (%): Dino 78.05±1.36 -> H-MGDM 82.06±1.17; GraphMAE2 72.56±1.80 -> H-MGDM 74.51±1.13; DiffMAE 84.58±1.72 -> H-MGDM 86.23±2.31 | Helped |
| use H-MGDM features instead of the strongest listed baseline | survival prediction · TCGA-KIRC, TCGA-ESCA, and CRC-PM | C-index: TCGA-KIRC DeepSurv: 67.92±5.61 vs 66.99±4.56; other listed backbone-dataset comparisons favor H-MGDM | Mixed |
Surprising findings
- Removing edge latents slightly improves CRC-PM survival C-index with DeepSurv (62.32 vs 62.27), contrary to the full model's advantage.
- Mask-ratio sweep peaks at intermediate ratios (0.6 for Komura, 0.7 for PANDA, 0.5 for IBD) rather than monotonically.
Limitations
- Relies on private datasets (IBD, CRC-PM), limiting reproducibility.
- Several gains fall within overlapping standard deviations; no formal significance testing beyond KM log-rank.
- Entity graph uses SLIC superpixels, which the authors suggest could be improved with newer entity extraction methods.
Experimental setup
Datasets: Komura et al. PANDA IBD TCGA-KIRC TCGA-ESCA CRC-PM
Benchmarks: Komura et al. PANDA IBD TCGA-KIRC TCGA-ESCA CRC-PM
Metrics: ACC F1 C-index RMSE
Closest papers by meaning
- Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction (NeurIPS 2026 · similarity 0.88)
- Learned Representation-Guided Diffusion Models for Large-Image Generation (CVPR 2024 · similarity 0.88)
- Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis (CVPR 2024 · similarity 0.88)
- STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation (NeurIPS 2026 · similarity 0.88)
- Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.88)
- Rotation-Agnostic Image Representation Learning for Digital Pathology (CVPR 2024 · similarity 0.86)
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.