Research map / Label-efficient and robust learning / Recognition and long-tail
Boosting Multiple Instance Learning Models for Whole Slide Image Classification: A Model-Agnostic Framework Based on Counterfactual Inference
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 CIMIL, a model-agnostic framework that boosts existing multiple instance learning models for WSI classification using counterfactual-inference sub-bag assessment, hierarchical instance searching for pseudo labels, and instance-embedding feature refinement.
Problem
Weakly supervised WSI classification with only bag-level labels, where existing MIL methods either focus on bag prediction or rely on model-specific mechanisms (e.g., attention) for accurate instance-level prediction.
Approach
Estimates the causal effect of masking sub-bags (replacing them with negative instances) on the bag prediction to score sub-bags, then recursively searches finer sub-bags to obtain reliable positive instance pseudo labels; the trained instance classifier's embeddings serve as prompts to refine instance features for any MIL backbone, without requiring attention.
Key results
- On CAMELYON16, CLAM-MB+Ours reaches 0.9015 bag AUC and 0.9429 instance AUC versus 0.8666 bag AUC for CLAM-MB alone.
- Ours applied to DSMIL reaches 0.9428 instance AUC on AGGC22, and MaxPooling+Ours reaches 0.9578 bag AUC on TCGA-NSCLC.
- Pseudo-label precision for positive patches with CLAM-MB rises from 0.4450 (baseline) to 0.8777 with the proposed method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CIMIL to CLAM-MB | CLAM-MB · CAMELYON16 bag (slide) | AUC: 0.8666 -> 0.9015 | Helped |
| Add CIMIL to CLAM-MB | CLAM-MB · CAMELYON16 instance (patch) | AUC: 0.8775 -> 0.9429 | Helped |
| Add CIMIL to CLAM-MB | CLAM-MB · AGGC22 instance (patch) | AUC: 0.7786 -> 0.8331 | Helped |
| Add CIMIL to CLAM-MB | CLAM-MB · TCGA-NSCLC bag (slide) | AUC: 0.9266 ± 0.0115 -> 0.9517 ± 0.0130 | Helped |
| Replace negative-instance masking with random-value masking (ablation) | CIMIL on CLAM-MB · CAMELYON16 bag (slide) | AUC: 0.9015 -> 0.8753 | Hurt |
| Replace negative-instance masking with an all-zero vector (ablation) | CIMIL on CLAM-MB · CAMELYON16 bag (slide) | AUC: 0.9015 -> 0.8863 | Hurt |
| Limit hierarchical instance searching to 1 layer (ablation) | CIMIL on CLAM-MB · CAMELYON16 instance (patch) | AUC: 0.9429 -> 0.9137 | Hurt |
| Limit hierarchical instance searching to 2 layers (ablation) | CIMIL on CLAM-MB · CAMELYON16 instance (patch) | AUC: 0.9429 -> 0.9372 | Hurt |
| Add a third-layer hierarchical instance search compared with a 2-layer limit (ablation) | CIMIL on CLAM-MB · CAMELYON16 instance (patch) | AUC: 0.9372 -> 0.9429 | Helped |
| Remove offline encoder feature F(xi) from feature refinement (ablation) | CIMIL on CLAM-MB · CAMELYON16 bag (slide) | AUC: 0.9015 -> 0.8809 | Hurt |
| Remove offline encoder feature F(xi) from feature refinement (ablation) | CIMIL on CLAM-MB · CAMELYON16 instance (patch) | AUC: 0.9429 -> 0.9429 | No clear effect |
Surprising findings
- Assigning slide labels directly to patches yields a pseudo-label precision of only 0.0318, showing that most patches in positive slides are negative.
Limitations
- Evaluated only on classification tasks; patch-level prediction is not tested as segmentation.
- Hierarchical search may discard some true positive patches.
- Boundary delineation of positive regions is coarse compared with ground truth.
Experimental setup
Datasets: CAMELYON16 AGGC22 TCGA-NSCLC
Benchmarks: CAMELYON16 AGGC22 TCGA-NSCLC
Metrics: AUC ACC balanced accuracy
Baselines: MaxPooling ABMIL DSMIL CLAM-SB CLAM-MB TransMIL DGMIL WENO
Builds on: ABMIL WENO DGMIL CLAM DSMIL
Closest papers by meaning
- OODML: Whole Slide Image Classification Meets Online Pseudo-Supervision and Dynamic Mutual Learning (AAAI 2025 · similarity 0.89)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 2025 · similarity 0.89)
- CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.88)
- ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.86)
- Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.86)
- Advancing Multiple Instance Learning with Continual Learning for Whole Slide Imaging (CVPR 2025 · 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.