Research map / Label-efficient and robust learning / Recognition and long-tail
CaMIL: Causal Multiple 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 CaMIL, a causal multiple instance learning framework that uses front-door adjustment with clustered instance-level features to estimate P(Y|do(X)) for whole slide image classification.
Problem
MIL-based WSI classification is prone to learning spurious associations (e.g., stain color correlated with disease) caused by unobserved confounders, which hurts generalization.
Approach
Replace the likelihood P(Y|X) with the interventional expectation P(Y|do(X)) via front-door adjustment, using the bag-level feature as mediator and approximating the marginalization over instances by K-means clusters of a buffer of instance features, fused by cross-attention and NWGM, in a single end-to-end training stage.
Key results
- On Camelyon16 with ResNet18, DTFD improves from 81.02 to 87.45 precision and AUC from 78.79 to 85.63 with the causal method.
- On TCGA-NSCLC with ViT-small, DTFD accuracy rises from 87.20 to 90.05 with the causal method, and ABMIL AUC from 93.87 to 94.36.
- Versus IBMIL on DSMIL backbone (Camelyon16, ResNet18), CaMIL reports precision 85.41 vs 81.99 for IBMIL, though IBMIL has higher AUC (83.85 vs 82.37).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CaMIL causal intervention to CLAM-SB | CLAM-SB with ViT-small · Camelyon16 | accuracy: +0.77 | Helped |
| Add CaMIL causal intervention to DSMIL | DSMIL with ViT-small · Camelyon16 | AUC: +2.59 | Helped |
| Add CaMIL causal intervention to TransMIL | TransMIL with ViT-small · TCGA-NSCLC | accuracy: +1.90 | Helped |
| Add CaMIL causal intervention to CLAM-SB | CLAM-SB with ViT-small · TCGA-NSCLC | AUC: +0.97 | Helped |
| Add CaMIL causal intervention to TransMIL | TransMIL with ViT-small · Camelyon16 | AUC: -0.27 | Hurt |
| Add CaMIL causal intervention to CLAM-SB | CLAM-SB with ViT-small · Camelyon16 | AUC: -0.02 | No clear effect |
| Add CaMIL causal intervention to DSMIL | DSMIL with ViT-small · TCGA-NSCLC | precision: -0.07 | Hurt |
| Add CaMIL causal intervention to TransMIL | TransMIL with ResNet18 · TCGA-NSCLC | AUC: -0.55 | Hurt |
| Vary the number of clusters to 2, 4, 8, or 16 (ablation) | CaMIL with ABMIL, max-pooling, or mean-pooling · TCGA-NSCLC | classification performance | Helped |
| Reduce the instances sampled per bag for the instance buffer (ablation) | CaMIL with ABMIL · TCGA-NSCLC | classification performance | Helped |
| Update the instance buffer less often (ablation) | CaMIL with ABMIL · TCGA-NSCLC | classification performance | Hurt |
| Increase the task-specific instance feature dimension (ablation) | CaMIL · TCGA-NSCLC | classification performance | Helped |
Surprising findings
- Ensembling multiple aggregators without causal intervention can yield worse results than the best single aggregator, while the causal method mitigates this.
- On Camelyon16 with ViT-small, the causal method slightly reduces AUC for TransMIL (-0.27) and DSMIL on TCGA (ViT) shows near-zero change.
Limitations
- Gains are sometimes small or negative on individual metrics (e.g., AUC drops for TransMIL and DSMIL with ViT-small on TCGA).
- Relies on approximations (NWGM and K-means clustering of a periodically updated instance buffer) rather than exact interventional expectation.
Experimental setup
Datasets: Camelyon16 TCGA-NSCLC
Benchmarks: Camelyon16 TCGA-NSCLC
Metrics: precision accuracy F1 AUC
Baselines: ABMIL DSMIL TransMIL CLAM-SB CLAM-MB DTFD IBMIL
Builds on: ABMIL DSMIL TransMIL CLAM DTFD IBMIL
Closest papers by meaning
- Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.90)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 2025 · similarity 0.90)
- Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction (CVPR 2024 · similarity 0.88)
- CAMIL: Context-Aware Multiple Instance Learning for Cancer Detection and Subtyping in Whole Slide Images (ICLR 2024 · similarity 0.88)
- ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.88)
- How Effective Can Dropout Be in Multiple Instance Learning ? (ICML 2025 · similarity 0.88)
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.