Research map / Label-efficient and robust learning / Recognition and long-tail
Interventional Bag Multi-Instance Learning on Whole-Slide Pathological Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IBMIL backdoor adjustment | TransMIL with CTransPath SRCL · Camelyon16 | AUC: 95.88 -> 97.00 | Helped |
| Add IBMIL backdoor adjustment | DSMIL with CTransPath SRCL · TCGA-NSCLC | AUC: 97.13 -> 97.51 | Helped |
| Freeze confounders instead of learning them end-to-end (ablation) | IBMIL with ResNet-18 + ABMIL · TCGA-NSCLC | accuracy: 83.81 -> 85.24 | Helped |
| Learn confounders end-to-end instead of freezing them (ablation) | IBMIL with ResNet-18 + ABMIL · TCGA-NSCLC | accuracy: 85.24 -> 83.81 | Hurt |
| Concatenate bag and confounder features rather than combine them by element-wise addition (ablation) | IBMIL with ResNet-18 + ABMIL · TCGA-NSCLC | accuracy: 84.76 -> 85.24 | Helped |
| Concatenate bag and confounder features rather than subtract them element-wise (ablation) | IBMIL with ResNet-18 + ABMIL · TCGA-NSCLC | accuracy: 84.29 -> 85.24 | Helped |
| Use non-parametric mean pooling to generate bag features instead of the trained aggregator (ablation) | IBMIL · TCGA-NSCLC | AUC: 80.68 -> 89.14 | Mixed |
| Use non-parametric max pooling to generate bag features instead of the trained aggregator (ablation) | IBMIL · TCGA-NSCLC | AUC: 82.23 -> 84.95 | Mixed |
| Vary the confounder dictionary size K (ablation) | IBMIL · TCGA-NSCLC | performance | No clear effect |
| Increase joint feature-space dimension (ablation) | IBMIL · TCGA-NSCLC | performance | Mixed |
| Use class-specific rather than class-agnostic K-means to construct confounders (ablation) | IBMIL · TCGA-NSCLC | performance | No clear effect |
| Train baseline methods for additional epochs instead of adding interventional training (ablation) | MIL baseline methods · WSI classification | performance | Hurt |
Closest papers by meaning
- CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.91)
- Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.89)
- ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.88)
- Boosting Multiple Instance Learning Models for Whole Slide Image Classification: A Model-Agnostic Framework Based on Counterfactual Inference (AAAI 2024 · similarity 0.88)
- Fast and Accurate Gigapixel Pathological Image Classification with Hierarchical Distillation Multi-Instance Learning (CVPR 2025 · similarity 0.87)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 2025 · 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.