MoEMambaMIL: Structure-Aware Selective State Space Modeling for Whole-Slide Image Analysis

Dongqing Xie; Yonghuang Wu

cs.CVMar 6, 2026

MoEMambaMIL: Structure-Aware Selective State Space Modeling for Whole-Slide Image Analysis

Authors: Dongqing Xie, Yonghuang Wu

Organizations: Tongji University · Fudan University

Abstract

Whole-slide image (WSI) analysis is challenging due to the gigapixel scale of slides and their inherent hierarchical multi-resolution structure. Existing multiple instance learning (MIL) approaches often model WSIs as unordered collections of patches, which limits their ability to capture structured dependencies between global tissue organization and local cellular patterns. Although recent State Space Models (SSMs) enable efficient modeling of long sequences, how to structure WSI tokens to fully exploit their spatial hierarchy remains an open problem.We propose MoEMambaMIL, a structure-aware SSM framework for WSI analysis that integrates region-nested selective scanning with mixture-of-experts (MoE) modeling. Leveraging multi-resolution preprocessing, MoEMambaMIL organizes patch tokens into region-aware sequences that preserve spatial containment across resolutions. On top of this structured sequence, we decouple resolution-aware encoding and region-adaptive contextual modeling via a combination of static, resolution-specific experts and dynamic sparse experts with learned routing. This design enables efficient long-sequence modeling while promoting expert specialization across heterogeneous diagnostic patterns. Experiments demonstrate that MoEMambaMIL achieves the best performance across 9 downstream tasks.

Figures & tables

Figure 1: Conceptual comparison between conventional MIL-based WSI modeling and the proposed MoEMambaMIL framework.

Figure 2: Overview of the proposed framework. Multi-resolution WSI patches are first organized into resolution-aware sequences and modeled by static experts to capture structural representations. A selective scan strategy is then applied to construct region-nested scans, yielding region-nested token sequences. These sequences are processed by a MoEMamba backbone with gating and routing mechanisms that dynamically dispatch tokens to Mamba experts. Finally, an attention-based MIL head aggregates token features for WSI-level prediction.

		TCGA Kidney			Liver Cancer			Camelyon 17
Methods	Feat.	F1	AUC	ACC	F1	AUC	ACC	F1	AUC	ACC
DSMIL †	RN	66.24	95.47	80.00	87.42	98.20	92.64	54.50	77.10	76.00
TransMIL †	RN	73.78	95.27	84.04	70.24	94.32	84.56	65.72	82.36	86.20
CLAM †	RN	64.75	94.84	79.55	84.24	98.43	92.12	63.52	77.03	87.00
CP-MID †	RN	84.21	97.47	88.88	86.65	97.98	92.46	62.66	78.37	86.40
CLAM †	UNI	90.53	99.14	92.13	83.72	98.36	92.96	73.97	85.57	87.00

Table 1: Performance comparison on three datasets with three feature extractors. RN: ResNet. Giga: GigaPath. † : from ( Wu et al., 2026 ) .

Figure 3: Performance differences (

\Delta=n-r

) between resolution-based (r) and region-nested (n) selective scanning across datasets and models. The mixed signs across metrics indicate complementary strengths of the two schemes.

Figure 4: Ablation study across different encoders (ResNet, UNI, Gigapath): Mean Performance Across Metrics.

TCGA Kidney
Methods	F1	AUC	ACC
WO/MoE	93.11	99.11	94.38
WO/R	91.62	99.08	93.26
MoEFFNMIL	92.88	99.15	93.82
MoEMambaMIL	95.16	99.43	96.63

Table 2: Ablation on each component. Dataset: Camelyon 17. Patch-level Encoder: Prov-Gigapath.

Layers	TopK	$\mathcal{L}_{\text{balance}}$	F1	AUC	ACC	MCC
2	2	0.001	82.98	91.52	95.00	86.76
3	2	0.001	83.16	92.50	95.00	86.85
4	2	0.001	87.00	91.70	95.00	87.19
5	2	0.001	88.86	92.11	96.00	89.15
6	2	0.001	89.99	95.24	96.00	90.00
6	1	0.001	84.57	97.76	95.00	86.94

Table 3: Ablation Study of MoE-Mamba Architecture Design.

Figure 5: Multi-resolution MIL attention visualization. For each WSI, visualizing instance-level attention weights derived from the final MIL pooling for three resolution levels (Level 0–2). Red indicates high attention, blue indicates low attention, and white denotes intermediate importance.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Description	Shape
$X$	Raw patch feature matrix from WSI	$[B,N,D_{\text{in}}]$
$r_{i}$	Resolution index of each patch	$[B,N]$
$H$	Patch features after embedding	$[B,N,D]$
$\tilde{H}$	Features after static expert encoding	$[B,N,D]$
$\mathcal{S}_{r}$	Index set of patches at resolution $r$	-
$f^{\text{static}}_{r}$	Static expert for resolution $r$	-

Appendix

Table 4: Symbol Mapping for Tensors and Functions

Hyperparameter	Symbol	Default	Description
Batch size	$B$	1	Number of WSIs per batch
Number of patches	$N$	varies	Patches per slide
Input feature dim	$D_{\text{in}}$	ResNet: 1024, UNI: 1024, Giga: 1536	Pretrained feature dimension
Hidden dimension	$D$	512	Model hidden dimension
Number of classes	$C$	task-specific	Classification categories
State dimension	$d_{\text{state}}$	16	Mamba state dimension

Appendix

Table 5: Hyperparameter Definitions and Default Settings

Aspect	Static Experts	Dynamic Experts
Routing	Deterministic (resolution-based)	Learned (content-based)
Scope	Resolution-specific	Shared across resolutions
Stage	Pre-scanning encoding	Post-scanning modeling
Purpose	Resolution-aware feature extraction	Region-adaptive pattern recognition

Appendix

Table 6: Complementary Roles of Static and Dynamic Experts.

Figure 6: Multi-resolution MIL attention visualization. For each WSI, visualizing instance-level attention weights derived from the final MIL pooling for three resolution levels (Level 0–2). Red indicates high attention, blue indicates low attention, and white denotes intermediate importance.

Explore similar work

KHiM-Mamba: Injecting Pathology Knowledge into Mamba via Hidden-State Modulation for Whole Slide Image Analysis
Aug 14, 2026Qixiang Zhang, Yi Li, Tianqi Xiang +4Whole-Slide Image Analysis State Space Models
Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong, China · Center for Intelligent Oncology, Chongqing University Cancer Hospital and Chongqing University School of Medicine, and Chongqing Key Laboratory of Intelligent Oncology for Breast Cancer, Chongqing, 400030, China
MambaBack: Bridging Local Features and Global Contexts in Whole Slide Image Analysis
Apr 17, 2026Sicheng Chen, Chad Wong, Tianyi Zhang +3Efficient Neural Network Inference Multiple Instance Learning
Nhu Department of Electrical Engineering and Computer Science, University of California, Irvine · Department of Electrical & Computer Engineering, National University of Singapore · PuzzleLogic Pte Ltd, Singapore 229594, Singapore
Geometry-Aware State Space Model: A New Paradigm for Whole-Slide Image Representation
May 6, 2026Enhui Chai, Sicheng Chen, Tianyi Zhang +4Histopathology Image Classification Multiple Instance Learning
PuzzleLogic Pte Ltd, Singapore 229594, Singapore · Nhu Department of Electrical Engineering and Computer Science, University of California, Irvine · Department of Electrical & Computer Engineering, National University of Singapore

Aug 14, 2026J/K move · Enter open · S save

KHiM-Mamba: Injecting Pathology Knowledge into Mamba via Hidden-State Modulation for Whole Slide Image Analysis

Qixiang Zhang, Yi Li, Tianqi Xiang, Haonan Wang+3

Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong, China · Center for Intelligent Oncology, Chongqing University Cancer Hospital and Chongqing University School of Medicine, and Chongqing Key Laboratory of Intelligent Oncology for Breast Cancer, Chongqing, 400030, China

Whole-Slide Image Analysis State Space Models Visual SSMs Multiple Instance Learning

Open paper AskPDF arXiv

Whole slide image analysis is commonly formulated as multiple instance learning (MIL), where instance features are contextually updated and aggregated into a slide representation, a process we term slide encoding dynamics. Recently, selective state-space models (SSM) have emerged as promising MIL architectures due to their long-sequence modeling capability and linear complexity. However, existing SSM-based MIL methods rely solely on visual features during MIL. Meanwhile, in large-scale WSIs, where sparse diagnostically decisive regions are surrounded by abundant irrelevant information, such purely vision-driven selective dynamics can misallocate state updates and readouts, causing the evolving SSM state to accumulate task-irrelevant evidence and dilute critical diagnostic cues over long scan trajectories. In this work, we propose the Knowledge-Aware Hidden-State Modulation architecture (KHiM-Mamba), which innovatively regulates Mamba's core selective state-space mechanism with explicit knowledge priors, steering slide encoding dynamics toward diagnostically meaningful evidence accumulation. Specifically, we redesign the original SSM layer to perform knowledge modulation operations during the evolution of hidden states, thereby guiding what visual evidence is accumulated and retrieved from the hidden state at each encoding step. Furthermore, we additionally introduce a local-adaptive vocabulary retrieval module that uses large language models to assign each patch fine-grained, tissue-specific semantic descriptions, enabling precise modulation across diverse tasks. Experiments on 11 public benchmarks across 4 tasks show that KHiM-Mamba consistently achieves state-of-the-art performance.

Abstract

Explore similar work

KHiM-Mamba: Injecting Pathology Knowledge into Mamba via Hidden-State Modulation for Whole Slide Image Analysis

MambaBack: Bridging Local Features and Global Contexts in Whole Slide Image Analysis

Geometry-Aware State Space Model: A New Paradigm for Whole-Slide Image Representation