Research map / Label-efficient and robust learning / Recognition and long-tail
Promptable Representation Distribution Learning and Data Augmentation for Gigapixel Histopathology WSI Analysis
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 PRDL, a DINO-based self-supervised framework that learns per-patch Gaussian representation distributions modulated by augmentation prompts, and PRS, which samples WSI-level training features from these distributions for promptable feature-space augmentation in MIL.
Problem
Patch representations in MIL-based WSI analysis are precomputed and fixed, making standard image-space data augmentation infeasible and feature-space augmentations either costly (generative models) or semantically lossy (mixup), with little control.
Approach
A distribution estimator predicts a Gaussian per patch during self-distillation; augmentation prompts (sampled operator compositions) narrow the variance through learned sigmoid masks so that sampling from the prompted distribution simulates specific image augmentations in feature space, enabling parameter-free online WSI augmentation at training time.
Key results
- On USTC-EGFR under the CLAM benchmark, PRDL+PRS reaches 89.4 AUC, 65.1 F1 and 65.9 ACC, versus 79.4 AUC for DINO features.
- On TCGA-EGFR under the CLAM benchmark, PRDL+PRS reaches 87.0 AUC versus 63.0 for DINO, and 83.4 AUC under TransMIL.
- PRDL+PRS achieves 97.5 to 97.8 AUC on TCGA-LUNG-3K across the three MIL frameworks, with the best AUC among the compared methods on that dataset under DTFD-MIL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add promptable representation augmentation (ablation) | DINO representation learning · USTC-EGFR, CLAM | AUC: 79.4 -> 86.4 | Helped |
| add promptable representation sampling (ablation) | PRDL · USTC-EGFR, CLAM | AUC: 86.4 -> 89.4 | Helped |
| remove KL divergence constraint (ablation) | PRDL · USTC-EGFR, CLAM | AUC: 86.4 -> 83.8 | Hurt |
| remove L1 mask sparsity constraint (ablation) | PRDL · USTC-EGFR, CLAM | AUC: 86.4 -> 83.8 | Hurt |
| remove mask variance regularization (ablation) | PRDL · USTC-EGFR, CLAM | AUC: 86.4 -> 84.6 | Hurt |
| use augmentation prompts rather than an unmasked broad augmentation range (ablation) | PRDL · USTC-EGFR, CLAM | AUC: 86.0 -> 89.4 | Helped |
| use random combinations of augmentation prompts instead of a single prompt (ablation) | PRDL+PRS · USTC-EGFR, CLAM | AUC: 88.8 -> 89.4 | Helped |
| add PRS to PRDL (ablation) | PRDL · TCGA-LUNG-3K, CLAM | AUC: 97.9 -> 97.5 | Hurt |
| use Mixup-based augmentation | DINO · TCGA-LUNG-3K, CLAM | AUC | Hurt |
| use standard Gaussian random perturbation | DINO · USTC-EGFR, CLAM | F1-Score: 14.8% | Hurt |
| use Monte Carlo sampling that randomly discards patch representations | DINO · USTC-EGFR, DTFD-MIL | AUC: 5.2% | Helped |
| use a single HorizontalFlip prompt rather than combined prompts (ablation) | PRDL · USTC-EGFR, CLAM | AUC: 89.4 -> 88.8 | Hurt |
Surprising findings
- Random Gaussian perturbation of representations lowered F1 under CLAM on USTC-EGFR relative to DINO, while Monte Carlo sampling hurt CLAM performance despite helping DTFD-MIL.
- Mixup-based augmentations (ReMix, RankMix) frequently degraded performance relative to DINO on TCGA-LUNG-3K under CLAM.
Limitations
- Gains are reported on three datasets, one of them private, with small test sets (e.g., 32 mutant WSIs in TCGA-EGFR).
- Hyperparameters were tuned on the USTC-EGFR validation subset and the method is built on a DINO baseline with a single ViT-S/16 backbone.
- Under the PAMA framework and foundation-model comparisons the advantage is mixed, e.g., CONCH is better on TCGA-LUNG-3K in some settings and PRDL alone is slightly below PRDL+PRS on some metrics.
Experimental setup
Datasets: USTC-EGFR TCGA-EGFR TCGA-LUNG-3K
Benchmarks: USTC-EGFR TCGA-EGFR TCGA-LUNG-3K
Metrics: AUC F1-score accuracy
Baselines: ImageNet SimCLR BYOL MoCov3 DINO DAGAN ReMix RankMix Intra-Mixup AugDiff CONCH PAMA
Builds on: DINO ABMIL CLAM TransMIL DTFD-MIL
Closest papers by meaning
- SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · similarity 0.89)
- CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.87)
- Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.86)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 2025 · similarity 0.86)
- MAPLE: Multi-scale Attribute-enhanced Prompt Learning for Few-shot Whole Slide Image Classification (NeurIPS 2025 · similarity 0.86)
- Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction (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.