Research map / Continual learning / Online and replay-based
Advancing Multiple Instance Learning with Continual Learning for Whole Slide Imaging
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.
Analyzes catastrophic forgetting in attention-based MIL for WSIs and proposes attention knowledge distillation (AKD) plus a pseudo-bag memory pool (PMP) for class-incremental continual learning.
Problem
Continual learning for multiple instance learning (MIL) on whole slide images suffers catastrophic forgetting, and replay is memory-expensive because each WSI contains tens of thousands of patches.
Approach
Decoupling and gradient analysis show that forgetting in attention MIL concentrates in the attention network rather than the classifier; AKD distills attention outputs on replayed samples, and PMP stores distilled pseudo-bags (MaxMinRand patch selection) instead of whole slides.
Key results
- On Camelyon-TCGA with CLAM and a 30-WSI memory, the method reaches 0.754 AACC versus 0.494 for ER with the same memory.
- On the Skin Cancer dataset, CLAM AACC improves from 0.5388 (ER) to 0.5926 and TransMIL reaches 0.6418 versus joint training 0.7647.
- Ablation on Camelyon-TCGA with CLAM: AKD alone gives 0.692 AACC and PMP alone 0.672, versus 0.296 for fine-tuning-level ER baseline and 0.729 for PMP+AKD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AKD instead of ER | CLAM · Skin Cancer | AACC: 0.5388 -> 0.5926 | Helped |
| Use AKD instead of MICIL w/ ER | TransMIL · Skin Cancer | AACC: 0.6275 -> 0.6418 | Helped |
| Add PMP and AKD | CLAM · Camelyon-TCGA, 5 WSIs | AACC: 0.301±0.010 -> 0.657±0.032 | Helped |
| Add PMP and AKD | CLAM · Camelyon-TCGA, 10 WSIs | AACC: 0.372±0.070 -> 0.729±0.041 | Helped |
| Add PMP and AKD | CLAM · Camelyon-TCGA, 30 WSIs | AACC: 0.494±0.058 -> 0.754±0.029 | Helped |
| Add PMP and AKD | TransMIL · Camelyon-TCGA, 30 WSIs | AACC: 0.315±0.028 -> 0.489±0.059 | Helped |
| Select random, highest-attention, and lowest-attention patches rather than only highest-attention patches (ablation) | Pseudo-bag memory pool for CLAM · Camelyon-TCGA | AACC: 0.595±0.121 -> 0.729±0.041 | Helped |
| Select random, highest-attention, and lowest-attention patches rather than only highest-attention patches (ablation) | Pseudo-bag memory pool for TransMIL · Camelyon-TCGA | AACC: 0.340±0.008 -> 0.394±0.085 | Helped |
| Add attention knowledge distillation without PMP (ablation) | CLAM · Camelyon-TCGA | AACC: 0.296±0.014 -> 0.692±0.037 | Helped |
| Add pseudo-bag memory pool without AKD (ablation) | CLAM · Camelyon-TCGA | AACC: 0.296±0.014 -> 0.672±0.025 | Helped |
| Add attention knowledge distillation without PMP (ablation) | TransMIL · Camelyon-TCGA | AACC: 0.290±0.002 -> 0.373±0.045 | Helped |
| Add pseudo-bag memory pool without AKD (ablation) | TransMIL · Camelyon-TCGA | AACC: 0.290±0.002 -> 0.348±0.052 | Helped |
Surprising findings
- Swapping classifiers across CL sessions barely hurts accuracy, while swapping attention networks causes large drops, indicating attention drift drives forgetting in MIL.
- Random and MaxMin patch selection for pseudo-bags perform comparably, suggesting AKD relies on the global attention distribution rather than only top-attention patches.
Limitations
- Still requires storing past patch features, which may pose privacy risks.
- Gains on TransMIL for Camelyon-TCGA are modest and variance is high (e.g., 0.394±0.085).
Experimental setup
Datasets: Skin Cancer dataset Camelyon16 TCGA-LUNG TCGA-RCC
Benchmarks: Skin Cancer dataset Camelyon-TCGA
Metrics: balanced average accuracy (AACC) backward transfer (BWT) intransigence measure (IM)
Baselines: Fine-tuning Joint training LwF MICIL ER DER++ MICIL w/ ER
Builds on: ABMIL CLAM TransMIL experience replay DER++ LwF
Closest papers by meaning
- ASMIL: Attention-Stabilized Multiple Instance Learning for Whole-Slide Imaging (ICLR 2026 · similarity 0.91)
- M3amba: Memory Mamba is All You Need for Whole Slide Image Classification (CVPR 2025 · similarity 0.87)
- Boosting Multiple Instance Learning Models for Whole Slide Image Classification: A Model-Agnostic Framework Based on Counterfactual Inference (AAAI 2024 · similarity 0.86)
- Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.86)
- ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.86)
- Continual Self-supervised Learning: Towards Universal Multi-modal Medical Data Representation Learning (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.