Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Heterogeneous Tissues with Mixture of Experts for Gigapixel Whole Slide Images
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 PAMoE, a plug-and-play mixture-of-experts layer with expert-choice routing and prototype-supervised experts that learns tissue heterogeneity in gigapixel WSI survival prediction without extra inference-time priors.
Problem
Gigapixel WSI analysis via MIL struggles to model intratumoral tissue heterogeneity and irrelevant patches; prior methods need pre-classification or clustering priors and specialized workflows, limiting scalability.
Approach
Combines expert-choice MoE routing (experts select top-k patches, unselected patches are dropped) with a subset of experts supervised by cosine-similarity prototypes of pathological tissue types (tumor, stroma, immune infiltration, necrosis) derived from CONCH classifications, while other experts remain free; prior is used only during training.
Key results
- PAMoE improved TransMIL C-index on COAD (0.676 to 0.688), LUAD (0.644 to 0.668), and PAAD (0.646 to 0.655) under 5-fold cross-validation.
- LongViT+PAMoE improved C-index on COAD from 0.655 to 0.689 and BRCA from 0.662 to 0.673.
- PatchGCN+PAMoE gave limited and inconsistent changes, e.g., BRCA dropped from 0.692 to 0.679.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PAMoE to TransMIL | TransMIL · COAD, LGG, LUAD, PAAD, BRCA survival prediction | C-index: 0.676 ± 0.028 -> 0.688 ± 0.036; 0.770 ± 0.062 -> 0.779 ± 0.072; 0.644 ± 0.026 -> 0.668 ± 0.016; 0.646 ± 0.026 -> 0.655 ± | Helped |
| Add PAMoE to LongViT | LongViT · COAD, LGG, LUAD, PAAD, BRCA survival prediction | C-index: 0.655 ± 0.043 -> 0.689 ± 0.034; 0.776 ± 0.071 -> 0.780 ± 0.060; 0.615 ± 0.053 -> 0.644 ± 0.048; 0.612 ± 0.079 -> 0.639 ± | Helped |
| Add PAMoE to PatchGCN | PatchGCN · COAD, LGG, LUAD, PAAD, BRCA survival prediction | C-index: 0.712 ± 0.033 -> 0.715 ± 0.036; 0.786 ± 0.062 -> 0.793 ± 0.055; 0.657 ± 0.016 -> 0.651 ± 0.047; 0.616 ± 0.065 -> 0.612 ± | Mixed |
| Add prior-supervised experts rather than using only free experts, at matched total expert count (ablation) | TransMIL+PAMoE · COAD, LGG, LUAD, PAAD, BRCA survival prediction; 4, 6, and 8 experts | C-index: COAD: 0.669 ± 0.043 -> 0.689 ± 0.045; 0.685 ± 0.015 -> 0.688 ± 0.036; 0.681 ± 0.035 -> 0.682 ± 0.041 | Mixed |
| Use 4 prior-supervised experts and 0 free experts instead of 4 prior-supervised and 2 free experts (ablation) | TransMIL+PAMoE · COAD, LGG, LUAD, PAAD, BRCA survival prediction | C-index: COAD: 0.688 ± 0.036 -> 0.689 ± 0.045; LGG: 0.779 ± 0.072 -> 0.775 ± 0.051; LUAD: 0.668 ± 0.016 -> 0.666 ± 0.033; PAAD: 0 | Mixed |
| Replace the PAMoE router with cosine-similarity assignment (CSA) (ablation) | TransMIL · COAD, LGG, LUAD, PAAD, BRCA survival prediction | C-index: COAD: 0.688 ± 0.036 -> 0.692 ± 0.039; LGG: 0.779 ± 0.072 -> 0.769 ± 0.063; LUAD: 0.668 ± 0.016 -> 0.665 ± 0.018; PAAD: 0 | Mixed |
| Add the prior-routing loss with non-zero α instead of using task loss alone (ablation) | TransMIL+PAMoE · COAD, LGG, LUAD, PAAD, BRCA survival prediction | C-index | Helped |
Surprising findings
- PAMoE gave limited and inconsistent gains on PatchGCN, a non-transformer model, and decreased its BRCA and LUAD scores.
- Cosine similarity assignment (fixed prior routing) is often competitive with PAMoE, performing lower than PAMoE on most but not all datasets.
Limitations
- Not evaluated on larger-scale models due to hardware constraints.
- Impact on non-transformer architectures is unclear; PatchGCN results are inconsistent.
- Gains on several datasets are within standard deviations of baselines.
Experimental setup
Datasets: TCGA COAD TCGA LGG TCGA LUAD TCGA PAAD TCGA BRCA
Benchmarks: TCGA COAD TCGA LGG TCGA LUAD TCGA PAAD TCGA BRCA
Metrics: C-index
Baselines: ABMIL AttnMISL CaMIL PANTHER HEAT PatchGCN TransMIL LongViT
Builds on: Mixture of Experts (Shazeer et al.) expert choice routing TransMIL LongViT PatchGCN PANTHER
Closest papers by meaning
- Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology (CVPR 2024 · similarity 0.87)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.86)
- CPath-Omni: A Unified Multimodal Foundation Model for Patch and Whole Slide Image Analysis in Computational Pathology (CVPR 2025 · similarity 0.86)
- Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning (CVPR 2025 · similarity 0.86)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.85)
- Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.85)
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.