Research map / Label-efficient and robust learning / Recognition and long-tail
OODML: Whole Slide Image Classification Meets Online Pseudo-Supervision and Dynamic Mutual 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 OODML, an online pseudo-supervision and dynamic mutual learning framework for weakly supervised WSI classification that jointly refines pseudo-bag features and pseudo-labels via an adaptive memory bank and decision-revision pseudo-labels.
Problem
Bag-label multi-instance learning for whole slide image classification yields unreliable pseudo-labels and poor instance/pseudo-bag feature representations, since offline or lagged pseudo-label generation and Top-K instance selection are used.
Approach
Online pseudo-labels for pseudo-bag tokens are generated by revising attention scores with bag-level predictions (DRPL), an adaptive memory bank stores the most label-related tokens for fusion into current tokens (SPFF), and a dynamic module fuses explicit and implicit bag predictions, with feature and pseudo-label learning optimized mutually.
Key results
- On CAMELYON16, OODML♠ reaches 0.978 accuracy and 0.987 AUC versus 0.945 accuracy for BCL, a 3.3% accuracy gain.
- On TCGA Lung, OODML♠ reaches 0.982 accuracy, versus 0.908 for BCL, with the paper reporting a 6.9% gain over the best pseudo-label method.
- On BRACS with ViT16SSL features, OODML♠ achieves 0.843 F1 and 0.960 AUC, outperforming ACMIL's 0.722 F1 and 0.888 AUC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Dynamic Decision-Making (DDM) weighting of explicit/implicit predictions instead of average (ablation) | OODML · BRACS (Res18ImageNet) | Accuracy: 0.770 to 0.837 | Helped |
| add DRPL pseudo-label generation (vs. none) (ablation) | w/o DRPL model · BRACS (Res18ImageNet) | Accuracy: about 5.8 points (0.712 to 0.770) | Helped |
| use DRPL pseudo-labels instead of BCE on bag label only (ablation) | w BCE baseline · BRACS (Res18ImageNet) | Accuracy: 0.574 to 0.640 | Helped |
| add multi-level feature fusion (MLFF) to the AMB pipeline (ablation) | OODML · BRACS (Res18ImageNet) | Accuracy: 0.770 vs 0.651 (w/o MLFF) | Helped |
| replace MLFF weighting with random/mixup fusion variants (ablation) | OODML · BRACS (Res18ImageNet) | Accuracy: 0.770 to 0.627 (mixup) and 0.667 (Wk=Fk/Ftop) | Hurt |
| set AMB size K to 5 vs larger values (10, 20) (ablation) | OODML⋆ · BRACS (Res18ImageNet) | Accuracy: K=5: 0.770; K=10: 0.720; K=20: 0.671 | Mixed |
| add DDM-style concatenation of explicit/implicit representations with refined features (ablation) | OODML · BRACS (Res18ImageNet) | AUC: AUC drops 1.2 points vs Concat(HIR;HER) | Hurt |
| full OODML (DDM-weighted) compared to ABMIL-style and pseudo-label baselines on CAMELYON16 | OODML · CAMELYON16 | Accuracy: 0.978 vs 0.945 (BCL) | Helped |
| full OODML on BRACS ViT16SSL features vs ACMIL | OODML · BRACS (ViT16SSL) | F1: 0.843 vs 0.722 (ACMIL) | Helped |
Limitations
- Performance relies on a memory mechanism whose design the authors identify as a limitation.
- Hyperparameters such as AMB size K and pseudo-bag size M are sensitive, with large values causing redundant interference.
Experimental setup
Datasets: CAMELYON16 TCGA Lung TCGA ESCA BRACS
Benchmarks: CAMELYON16 TCGA Lung TCGA ESCA BRACS
Metrics: accuracy F1 AUC precision recall
Baselines: ABMIL MeanPooling MaxPooling DSMIL CLAM TransMIL DTFD MHIM MuRCL BCL CIMIL ACMIL
Builds on: ABMIL DTFD CLAM TransMIL BCL
Closest papers by meaning
- Boosting Multiple Instance Learning Models for Whole Slide Image Classification: A Model-Agnostic Framework Based on Counterfactual Inference (AAAI 2024 · similarity 0.89)
- FAST: A Dual-tier Few-Shot Learning Paradigm for Whole Slide Image Classification (NeurIPS 2024 · similarity 0.87)
- Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.86)
- MExD: An Expert-Infused Diffusion Model for Whole-Slide Image Classification (CVPR 2025 · similarity 0.85)
- Navigating Real-World Partial Label Learning: Unveiling Fine-Grained Images with Attributes (AAAI 2024 · similarity 0.85)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 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.