Research map / Detection and segmentation / CNN features
CAMIL: Context-Aware Multiple Instance Learning for Cancer Detection and Subtyping in Whole Slide Images
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 transformer-based MIL model whose attention is constrained by a neighbor-similarity mask over WSI tiles, to improve WSI cancer detection and subtyping.
Problem
Attention-based MIL models for whole slide image classification ignore contextual and spatial relationships among neighboring tiles, causing misclassification of sparse tumor regions and noisy instances.
Approach
Builds a neighbor-constrained attention module in which each tile's attention is recalibrated by summing its 8-neighbors' query-key scores weighted by a feature-similarity mask from a frozen SimCLR encoder, and combines it with a Nystromformer global transformer branch and gated local/global fusion.
Key results
- CAMIL achieves test AUC 0.959 on CAMELYON16, 0.975 on TCGA-NSCLC, and 0.881 on CAMELYON17 (ties DTFD-MIL's 0.884 on CAMELYON17 is higher; CAMIL trails by 0.003).
- CAMIL reaches ACC 0.917 on CAMELYON16 and 0.843 on CAMELYON17, the best among compared MIL methods.
- Ablation: CAMIL-L (neighbor attention only) reaches AUC 0.953 on CAMELYON16 vs 0.950 for CAMIL-G (Nystromformer only); full CAMIL is best at 0.959.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add neighbor-constrained attention to CAMIL-G (ablation) | CAMIL · CAMELYON16 | ACC: 0.891 -> 0.917 | Helped |
| add neighbor-constrained attention to CAMIL-G (ablation) | CAMIL · CAMELYON16 | AUC: 0.950 -> 0.959 | Helped |
| add Nystromformer to CAMIL-L (ablation) | CAMIL · CAMELYON16 | ACC: 0.910 -> 0.917 | Helped |
| add Nystromformer to CAMIL-L (ablation) | CAMIL · CAMELYON16 | AUC: 0.953 -> 0.959 | Helped |
| add neighbor-constrained attention to CAMIL-G (ablation) | CAMIL · TCGA-NSCLC | ACC: 0.907 -> 0.916 | Helped |
| add neighbor-constrained attention to CAMIL-G (ablation) | CAMIL · TCGA-NSCLC | AUC: 0.973 -> 0.975 | Helped |
| add Nystromformer to CAMIL-L (ablation) | CAMIL · TCGA-NSCLC | ACC: 0.914 -> 0.916 | Helped |
| add Nystromformer to CAMIL-L (ablation) | CAMIL · TCGA-NSCLC | AUC: 0.975 -> 0.975 | No clear effect |
| add neighbor-constrained attention to CAMIL-G (ablation) | CAMIL · CAMELYON17 | ACC: 0.818 -> 0.843 | Helped |
| add neighbor-constrained attention to CAMIL-G (ablation) | CAMIL · CAMELYON17 | AUC: 0.875 -> 0.881 | Helped |
| add Nystromformer to CAMIL-L (ablation) | CAMIL · CAMELYON17 | ACC: 0.828 -> 0.843 | Helped |
| replace ImageNet-pretrained ResNet-18 with SimCLR-pretrained feature extractor (ablation) | CAMIL · CAMELYON16 | ACC | Helped |
Surprising findings
- Using ImageNet-pretrained ResNet-18 features instead of SimCLR features drops CAMELYON16 ACC to 0.723 and AUC to 0.743, showing dependence on the feature extractor.
- CAMIL's localization Dice (0.515) is slightly below DTFD-MIL (0.525), which the authors attribute to the Nystromformer module.
Limitations
- Localization Dice is slightly behind DTFD-MIL, attributed to the Nystromformer module.
- Performance depends heavily on SimCLR-pretrained features; ImageNet features degrade results substantially.
- CAMELYON17 AUC does not beat DTFD-MIL; GTP underperforms on CAMELYON datasets.
Experimental setup
Datasets: CAMELYON16 CAMELYON17 TCGA-NSCLC
Benchmarks: CAMELYON16 TCGA-NSCLC CAMELYON17
Metrics: AUC accuracy F1 Dice specificity
Baselines: CLAM-SB CLAM-MB TransMIL DTFD-MIL DSMIL GTP
Builds on: attention-based MIL Nystromformer SimCLR TransMIL CLAM
Closest papers by meaning
- CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.88)
- Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.86)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 2025 · similarity 0.86)
- SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · similarity 0.85)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.85)
- ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification (CVPR 2024 · 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.