Research map / Label-efficient and robust learning / Recognition and long-tail
Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification
Yunlong Zhang, Honglin Li, YUXUAN SUN, Chenglu Zhu, Sunyi Zheng, Lin Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 diversity loss L_d to multiple branch attention (ablation) | ACMIL · CAMELYON16 (ViT-S/16 SSL) | AUC: 0.943 -> 0.974 | Improved |
| add diversity loss L_d to multiple branch attention (ablation) | ACMIL · BRACS (ResNet-18) | AUC: 0.760 -> 0.754 | No clear effect |
| keep STKIM active at test time (ablation) | ACMIL · CAMELYON16/BRACS/LBC (ViT-S/16 SSL) | F1/AUC: 0.954 -> 0.927 F1 on CAMELYON16 | Hurt |
| mask all Top-K instances (p=1.0) in STKIM (ablation) | ACMIL · CAMELYON16/BRACS/LBC | AUC | Hurt |
| use WENO-style masking (95 instances, p=1.0) instead of STKIM default (ablation) | ACMIL · CAMELYON16 (ViT-S/16 SSL) | AUC: 0.974 -> 0.843 | Hurt |
| vary number of masked instances K in STKIM (ablation) | ACMIL · CAMELYON16/BRACS/LBC | AUC: less than 1.0% AUC difference | No clear effect |
| add Multiple Branch Attention to ABMIL (ablation) | ABMIL · CAMELYON16 (ResNet-18) | AUC: 0.790 -> 0.850 | Improved |
| add STKIM to ABMIL (ablation) | ABMIL · BRACS (ResNet-18) | AUC: 0.723 -> 0.789 | Improved |
| combine MBA with STKIM (ablation) | ACMIL · CAMELYON16/BRACS/LBC | AUC: 39 of 45 configurations exceed MBA/STKIM-alone line | Improved |
| apply ACMIL on top of multi-head attention instead of gated attention | MHA baseline · BRACS (ResNet-18) | AUC: 0.738 -> 0.723 | Mixed |
| apply STKIM masking at training, versus MHIM-MIL teacher-student masking | ABMIL · CAMELYON16 (ResNet-18) | training time per epoch: 20.8s -> 8.0s | Improved |
| add STKIM | ABMIL attention · CAMELYON16 | top-10 cumulative attention: 0.87 -> 0.6 | Improved |
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 0.90)
- Advancing Multiple Instance Learning with Continual Learning for Whole Slide Imaging (CVPR 2025 · similarity 0.88)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 2025 · similarity 0.87)
- CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.86)
- How Effective Can Dropout Be in Multiple Instance Learning ? (ICML 2025 · similarity 0.86)
- ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification (CVPR 2024 · 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.
