Research map / Label-efficient and robust learning / Recognition and long-tail
Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification
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 PAMIL, a WSI classification framework that couples RL-based dynamic instance sampling (DPIS, trained with PPO and reward/penalty feedback) with a selection fusion feature representation (SFFR) and a Transformer classifier, evaluated on CAMELYON16 and TCGA Lung.
Problem
Weakly supervised whole slide image classification under multi-instance learning suffers from costly instance sampling, over-fitting in feature representation, and pseudo-bag bias affecting decision-making.
Approach
Treats instance sampling as an MDP solved with PPO, using feature-distance and neighbor-relation based policies plus reward/punishment signals from the Transformer classifier, and fuses past sampled tokens via MHA and a Siamese contrastive loss to aggregate historical sub-bag information.
Key results
- On CAMELYON16 DPIS-LIIS reaches 0.961 accuracy and 0.977 AUC, versus 0.925 accuracy for MHIM (DSMIL) (about 3.8% gain claimed).
- On TCGA Lung DPIS-LIIS reaches 0.965 accuracy and 0.994 AUC, outperforming IAT by about 4.4% accuracy claimed.
- Adding DPSF module to MaxPooling raises CAMELYON16 accuracy from 0.826 to 0.930 and AUC from 0.854 to 0.938.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPIS-LIIS instead of the strongest listed baseline | WSI classification · CAMELYON16 | accuracy: 0.925 -> 0.961 | Helped |
| Use DPIS-LIIS instead of the strongest listed baseline | WSI classification · TCGA Lung | accuracy: 0.921 -> 0.965 | Helped |
| Use pseudo-bag size 512 rather than the baseline size 512 random-sampling setup alternatives (ablation) | PAMIL on CAMELYON16 · CAMELYON16 | accuracy: 0.923 (128); 0.923 (256); 0.954 (512); 0.930 (1024); 0.907 (2048) | Mixed |
| Replace dynamic sampling with fixed K-Means grouping (ablation) | PAMIL on CAMELYON16 · CAMELYON16 | accuracy: 0.954 -> 0.907 (5 groups); 0.899 (10 groups); 0.915 (15 groups) | Hurt |
| Replace dynamic sampling with random grouping (ablation) | PAMIL on CAMELYON16 · CAMELYON16 | accuracy: 0.954 -> 0.884 (5 groups); 0.920 (10 groups); 0.907 (15 groups) | Hurt |
| Add attention without MHA fusion (ablation) | PAMIL feature representation · CAMELYON16 | accuracy: 0.938 -> 0.946 | Helped |
| Add MHA fusion without attention (ablation) | PAMIL feature representation · CAMELYON16 | accuracy: 0.938 -> 0.954 | Helped |
| Add SIA loss (ablation) | PAMIL optimization · CAMELYON16 | accuracy: 0.915 -> 0.930 (without STL); 0.915 -> 0.954 (with STL) | Helped |
| Add STL loss (ablation) | PAMIL optimization · CAMELYON16 | accuracy: 0.915 -> 0.915 (without SIA); 0.930 -> 0.954 (with SIA) | Mixed |
| Add reward feedback (ablation) | PAMIL DPIS · CAMELYON16 | accuracy: 0.930 -> 0.915 (without penalty); 0.938 -> 0.954 (with penalty) | Mixed |
| Add penalty feedback (ablation) | PAMIL DPIS · CAMELYON16 | accuracy: 0.930 -> 0.938 (without reward); 0.915 -> 0.954 (with reward) | Helped |
Surprising findings
- Random grouping into 10 groups (accuracy 0.920) outperforms K-Means grouping (0.899) and position sorting-based sampling underperforms on CAMELYON16 despite strong TCGA results.
Limitations
- Distance-guided sampling does not fully exploit the complex structural information of WSIs.
- Pseudo-label sensitivity can introduce bias, as the authors note.
- Results on CAMELYON16 show high variance for some variants (e.g., DPIS-GMSS 0.963±0.033).
Experimental setup
Datasets: CAMELYON16 TCGA Lung Cancer
Benchmarks: CAMELYON16 TCGA Lung Cancer
Metrics: AUC accuracy F1 precision recall
Baselines: MeanPooling MaxPooling ABMIL RNNMIL DSMIL CLAM TransMIL DTFD DGMIL IAT MHIM
Builds on: multi-instance learning proximal policy optimization DTFD MHIM TransMIL
Closest papers by meaning
- CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.90)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 2025 · similarity 0.89)
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.87)
- Transformer-Based Video-Structure Multi-Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.87)
- Fast and Accurate Gigapixel Pathological Image Classification with Hierarchical Distillation Multi-Instance Learning (CVPR 2025 · similarity 0.87)
- Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis (CVPR 2024 · similarity 0.87)
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.