Research map / Label-efficient and robust learning / Recognition and long-tail
Do Multiple Instance Learning Models Transfer?
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.
Systematically evaluates transfer of 11 MIL architectures across 21 pretraining tasks in computational pathology, finding that supervised pretraining (especially on the pancancer PC-108 task) consistently beats random initialization and matches or beats slide foundation models with far less pretraining data.
Problem
Random initialization is the standard for MIL models in computational pathology, and little is known about whether supervised slide-level MIL models transfer to new tasks, which matters for small weakly-supervised clinical datasets.
Approach
Treat supervised MIL models trained on large hierarchical pancancer classification (PC-108, PC-43) as transferable slide-level foundation models, and systematically measure transfer via KNN and end-to-end finetuning across architectures, scales, encoders, and pretraining tasks.
Key results
- Pretraining on PC-108 improves average finetuning performance over random initialization by 3.3% across 11 MIL architectures, with ABMIL at top.
- PC-108 frozen KNN features beat CHIEF in 12/15 tasks (+5.9% average) and GigaPath in 13/15 tasks (+9.7% average), using a pretraining set 6.5% and 2.3% the size of theirs.
- Few-shot experiments show DFTD gains a 171% increase at K=4 shots over random initialization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pretrain MIL models on PC-108 before downstream finetuning | MIL models · 19 CPath tasks | average performance: 3.3% | Helped |
| Use pancancer PC-108 rather than random initialization | ABMIL · BRACS | balanced accuracy: 49.7 -> 52.5 | Helped |
| Use PC-108 pretraining rather than random initialization | ABMIL · BCNB | AUROC: 80.9 -> 84.5 | Helped |
| Use PC-108 pretraining rather than random initialization | ABMIL · BRCA | AUROC: 69.2 -> 73.3 | Helped |
| Pretrain on PC-108 rather than random initialization | ABMIL · NSCLC Morph | AUROC: 95.3 -> 96.1 | Helped |
| Pretrain on PC-108 rather than random initialization | ABMIL · NSCLC Molec | AUROC: 67.4 -> 72.8 | Helped |
| Pretrain on PC-108 rather than random initialization | ABMIL · PANDA | weighted kappa: 91.6 -> 93.3 | Helped |
| Use PC-108 rather than random initialization | ABMIL with ResNet-50 features · 15 CPath tasks | average performance: 56.2 -> 59.6 | Helped |
| Use PC-108 rather than random initialization | ABMIL with UNIv2 features · 15 CPath tasks | average performance: 72.9 -> 74.6 | Helped |
| Use PC-108 rather than random initialization | ABMIL with CONCHv1.5 features · 15 CPath tasks | average performance: 72.3 -> 74.6 | Helped |
| Increase ABMIL size to 5.2M parameters and use PC-108 pretraining | ABMIL · 8 tasks | average performance: 70.7 -> 75.7 | Helped |
| Compare PC-108 ABMIL transfer with foundation models | ABMIL slide representations · 15 tasks, KNN | average performance: 60.8 -> 70.3; 54.5 -> 66.2 | Helped |
Surprising findings
- Pretraining on out-of-organ tasks (e.g., NSCLC to breast biopsy) transfers as well as in-domain pretraining.
- Reinitializing only the attention layer of a pretrained ABMIL causes the largest single performance drop, diverging from CNN transfer findings that later layers matter little.
Limitations
- Study excludes state-space MIL models and survival prediction tasks.
- PC-108 pretraining data come only from a single institution (Brigham and Women's Hospital), which may introduce bias.
- Gains vary across architectures; some tasks and models (e.g., DSMIL, CLAM) show small or negative transfer effects.
Experimental setup
Datasets: PC-108 PC-43 EBRAINS NSCLC (TCGA, CPTAC, NLST) PANDA BRACS TCGA-BRCA BCNB GBMLGG TCGA lung biomarkers
Benchmarks: EBRAINS NSCLC subtyping PANDA BRACS BCNB BRCA biomarkers GBMLGG NSCLC biomarkers
Metrics: AUROC balanced accuracy weighted kappa KNN accuracy
Baselines: random initialization CHIEF GigaPath DSMIL CLAM TransMIL Transformer DFTD RRT WIKG ILRA
Builds on: ABMIL CLAM TransMIL UNI CHIEF GigaPath
Closest papers by meaning
- Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology (CVPR 2024 · similarity 0.89)
- xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology (NeurIPS 2024 · similarity 0.88)
- Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology (NeurIPS 2025 · similarity 0.87)
- Symb-xMIL: Symbolic Explanations for Multiple Instance Learning in Digital Pathology (NeurIPS 2026 · similarity 0.87)
- SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · similarity 0.87)
- How Well Do Supervised 3D Models Transfer to Medical Imaging Tasks? (ICLR 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.