Organizations: Centre for Human Genetics, University of Oxford, Oxford, UK · Beckman Coulter Diagnostics, Brea, CA, USA · Danaher Corporation, Washington, DC, USA
Multiple Instance Learning (MIL) is the dominant framework for gigapixel whole-slide image (WSI) classification in computational pathology. However, current MIL aggregators route all instances through a shared pathway, constraining their capacity to specialise across the pathological heterogeneity inherent in each slide. Mixture-of-Experts (MoE) methods offer a natural remedy by partitioning instances across specialised expert subnetworks; yet unconstrained softmax routing may yield highly imbalanced utilisation, where one or a few experts absorb most routing mass, collapsing the mixture back to a near-single-pathway solution. To address these limitations, we propose ROAM (Region-graph OptimAl-transport Mixture-of-experts), a spatially aware MoE-MIL aggregator that routes region tokens to expert poolers via capacity-constrained entropic optimal transport, promoting balanced expert utilisation by construction. ROAM operates on spatial region tokens, obtained by compressing dense patch bags into spatially binned units that align routing with local tissue neighbourhoods and introduces two key mechanisms: (i) region-to-expert assignment formulated as entropic optimal transport (Sinkhorn) with explicit per slide capacity marginals, enforcing balanced expert utilisation without auxiliary load-balancing losses; and (ii) graph-regularised Sinkhorn iterations that diffuse routing assignments over the spatial region graph, encouraging neighbouring regions to coherently route to the same experts. Evaluated on four WSI benchmarks with frozen foundation-model patch embeddings, ROAM achieves performance competitive against strong MIL and MoE baselines, and on NSCLC generalisation (TCGA-CPTAC) reaches external AUC 0.845 +- 0.019.
Figures & tables
Figure 1: Frozen patch embeddings are pooled into M spatial region tokens. A routing GNN on the region graph parameterises region-to-expert costs, and Sinkhorn optimal transport routes region mass to E experts under per-slide capacity constraints (with optional graph smoothing). Each expert performs gated-attention pooling over its routed regions, and expert outputs are fused for slide-level prediction.
Enc.
Method
NSCLC (LUAD/LUSC)
BRCA
CRC
PANDA
int. AUC ↑
ext. AUC ↑
int. AUC ↑
int. AUC ↑
int. QWK ↑
UNI2-h
OTSurv [ 15 ]
0.964 ± 0.014
0.786 ± 0.021
0.884 ± 0.017
0.661 ± 0.055
0.873 ± 0.019
MeanPool
0.965 ± 0.013
0.800 ± 0.018
0.881 ± 0.023
0.671 ± 0.046
0.893 ± 0.012
MaxPool
0.964 ± 0.017
0.823 ± 0.021
0.867 ± 0.023
0.636 ± 0.046
0.811 ± 0.015
ILRA [ 22 ]
0.974 ± 0.012
0.827 ± 0.023
0.898 ± 0.021
0.702 ± 0.022
0.916 ± 0.005
CLAM-MB [ 13 ]
0.969 ± 0.020
0.831 ± 0.017
0.901 ± 0.021
0.697 ± 0.050
0.911 ± 0.009
Table 1: WSI benchmark results with UNI2 patch encoders (mean ± std over 5-folds). For NSCLC (LUAD/LUSC) , we report internal AUC on TCGA ( int. ) and external AUC on CPTAC ( ext. ); for BRCA and CRC we report internal AUC; for PANDA we report internal QWK. Best per column in bold ; second best underlined .
Figure 2: Qualitative routing visualisation on CPTAC (NSCLC). Two CPTAC external-test slides (top: LUAD; bottom: LUSC); numbers denote P(LUSC) . Left: ROAM dominant expert per region (territories). Middle/Right: ABMIL and CLAM-SB attention heatmaps (normalised for visualisation).
Configurations
int. AUC ↑
ext. AUC ↑
I
w/o routing GNN
0.971 ± 0.022
0.836 ± 0.023
II
w/o graph regularisation
0.966 ± 0.019
0.813 ± 0.035
III
softmax routing (no capacity constraint)
0.964 ± 0.029
0.809 ± 0.045
IV
w/o OT-guided pooling (no OT modulation)
0.966 ± 0.026
0.825 ± 0.015
V
OT-guided pooling (detach routing weights)
0.970 ± 0.019
0.820 ± 0.055
ROAM
0.976 ± 0.015
0.845 ± 0.019
Table 2: Ablation study on NSCLC (LUAD vs. LUSC) using UNI2-h embeddings (mean ± std over 5 folds). Internal evaluation is on TCGA ( int. ); external evaluation is on CPTAC ( ext. ).
Multiple Instance Learning (MIL) is the predominant framework for classifying gigapixel whole-slide images in computational pathology. MIL follows a sequence of 1) extracting patch features, 2) applying a linear layer to obtain task-specific patch features, and 3) aggregating the patches into a slide feature for classification. While substantial efforts have been devoted to optimizing patch feature extraction and aggregation, none have yet addressed the second point, the critical layer which transforms general-purpose features into task-specific features. We hypothesize that this layer constitutes an overlooked performance bottleneck and that stronger representations can be achieved with a low-rank transformation tailored to each patch's phenotype, yielding synergistic effects with any of the existing MIL approaches. To this end, we introduce MAMMOTH, a parameter-efficient, multi-head mixture of experts module designed to improve the performance of any MIL model with minimal alterations to the total number of parameters. Across eight MIL methods and 19 different classification tasks, we find that such task-specific transformation has a larger effect on performance than the choice of aggregation method. For instance, when equipped with MAMMOTH, even simple methods such as max or mean pooling attain higher average performance than any method with the standard linear layer. Overall, MAMMOTH improves performance in 130 of the 152 examined configurations, with an average +3.8% change in performance. Code is available at https://github.com/mahmoodlab/mammoth.
Daniel Shao, Joel Runevic, Richard J. Chen +4
Massachusetts Institute of Technology, Cambridge MA · Harvard University, Cambridge MA · Emory University, Atlanta GA +1
Whole slide image (WSI) analysis in computational pathology follows a multiple instance learning (MIL) pipeline where patch embeddings are extracted independently and aggregated for slide-level prediction, but within-slide variance from staining, scanner, and local texture can overwhelm the discriminative signal. We propose Masked Feature Encoding for Multiple Instance Learning (MFE-MIL), a feature-space masking framework that trains a lightweight MLP adapter jointly with a window-based masked reconstruction branch and a MIL classification head. The two objectives are complementary. Classification guides the adapter to suppress within-slide patch variance, while window-based masked reconstruction provides an auxiliary regularizer for the adapted features without using patch coordinates, coordinate graphs, or segmentation preprocessing. The raster patch-extraction order is used only as a weak implicit prior. At inference, the decoder is removed, leaving only the adapter and MIL head. Across CAMELYON16/17, PANDA, and TCGA-BRCA with four diverse encoders, MFE-MIL improves ACC/F1 for nearly all tested aggregator-encoder settings and AUC in most, outperforms coordinate-based spatial methods (CAMIL), and achieves higher AUC than 2DMamba on three of four datasets (UNI). On five TCGA survival cohorts it improves the average concordance index for every aggregator tested, its most consistent gain. Code is available at https://github.com/AtlasAnalyticsLab/MFE-MIL.
Haoyu He, Basile Tessier-Cloutier, Yang Wang +1
Department of Computer Science and Software Engineering (CSSE), Concordia University, Montreal, Canada · McGill University, Montreal, Canada · Mila – Quebec AI Institute, Montreal, Canada
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.