Research map / Architectures and efficiency / Attention and transformers
M3amba: Memory Mamba is All You Need for Whole Slide Image Classification
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 M3amba, a memory-driven grouped bidirectional Mamba with a dynamic memory bank and mixture-of-experts bag head for efficient whole slide image classification under multi-instance learning.
Problem
Whole slide image (WSI) classification with multi-instance learning needs long-range instance dependency modeling, but Transformers have quadratic cost and vanilla Mamba suffers from contextual/long-range forgetting and weak local modeling.
Approach
Randomly sample instances into groups, model intra-group relations with a BiMamba block, and store/refine historical group features in a dynamic memory bank accessed via cross-attention; then fuse inter-group BiMamba and convolutional global/local bag representations with a mixture-of-experts head.
Key results
- On TCGA BRCA, M3amba reaches 0.931 accuracy and 0.971 AUC versus 0.869 accuracy for the best prior method ACMIL (6.2% accuracy gain).
- On TCGA Lung, M3amba reaches 0.937 accuracy, a 7.0% gain over the strongest baseline (MHIM-ABMIL at 0.867).
- On BRACS with ResNet18-ImageNet features, M3amba reaches 0.800 F1 and 0.939 AUC, versus 0.596 F1 and 0.811 AUC for MambaMIL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use M3amba instead of ACMIL | TCGA BRCA classification · TCGA BRCA | accuracy: 6.2% | Helped |
| Use M3amba instead of ACMIL | TCGA Lung classification · TCGA Lung | accuracy: 7.0% | Helped |
| Use M3amba instead of ACMIL | TCGA ESCA classification · TCGA ESCA | accuracy: 3.8% | Helped |
| Add BiMamba to the Mamba baseline (ablation) | M3amba · BRACS | accuracy: 0.540 -> 0.586 | Helped |
| Add ConvBlock after BiMamba (ablation) | M3amba · BRACS | accuracy: 0.586 -> 0.609 | Helped |
| Add group cross-entropy loss (ablation) | M3amba · BRACS | accuracy: 0.609 -> 0.644 | Helped |
| Add DMB (ablation) | M3amba · BRACS | accuracy: 0.644 -> 0.724 | Helped |
| Add MoE (ablation) | M3amba · BRACS | accuracy: 0.724 -> 0.805 | Helped |
| Remove augmented DMB features from cross-attention (ablation) | M3amba · BRACS | accuracy: 0.724 -> 0.667 | Hurt |
| Replace DMB feature aggregation with random mixup (ablation) | M3amba · BRACS | precision: 0.700 vs 0.725 | Hurt |
| Set DMB size K to 3 instead of 0 (ablation) | M3amba · BRACS | AUC: 0.788 -> 0.888 | Helped |
| Increase DMB size from K=3 to K=10 or K=20 (ablation) | M3amba · BRACS | accuracy: 0.667 -> 0.609; 0.667 -> 0.598 | Hurt |
Limitations
- Evaluation relies on frozen ImageNet/SSL patch features rather than end-to-end training.
- Results on TCGA ESCA have high variance across folds (std up to 0.026-0.078 for baselines) with a small dataset of 156 slides.
Experimental setup
Datasets: TCGA BRCA TCGA Lung TCGA ESCA BRACS
Benchmarks: TCGA BRCA TCGA Lung TCGA ESCA BRACS
Metrics: accuracy AUC precision recall F1
Baselines: ABMIL DSMIL CLAM TransMIL DTFD MHIM-ABMIL MHIM-TransMIL IBMIL-ABMIL SSMMIL MambaMIL ACMIL
Builds on: Mamba MambaMIL TransMIL DTFD CLAM
Closest papers by meaning
- 2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification (CVPR 2025 · similarity 0.88)
- Advancing Multiple Instance Learning with Continual Learning for Whole Slide Imaging (CVPR 2025 · similarity 0.87)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.86)
- Mamba-Adaptor: State Space Model Adaptor for Visual Recognition (CVPR 2025 · similarity 0.85)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.85)
- GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.84)
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.